python

Python/OAK/datasets/_12_check_percent_status.py

@@ -0,0 +1,147 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+import json
+from pathlib import Path
+from collections import Counter, defaultdict
+import argparse
+
+
+def analisar_dataset(root: Path):
+    resultado = {}
+
+    for grupo_dir in sorted([p for p in root.iterdir() if p.is_dir()]):
+        labels_dir = grupo_dir / "labels"
+        if not labels_dir.is_dir():
+            continue
+
+        contagem = Counter()
+        total = 0
+
+        for json_path in sorted(labels_dir.glob("*.json")):
+            try:
+                with json_path.open("r", encoding="utf-8") as f:
+                    data = json.load(f)
+
+                status = data.get("estado_corredor")
+                if not status:
+                    status = "SEM_STATUS"
+
+                contagem[status] += 1
+                total += 1
+
+            except Exception:
+                contagem["ERRO_JSON"] += 1
+                total += 1
+
+        resultado[grupo_dir.name] = {
+            "total": total,
+            "contagem": contagem
+        }
+
+    return resultado
+
+
+def imprimir_relatorio(resultado):
+    todos_status = sorted({
+        status
+        for dados in resultado.values()
+        for status in dados["contagem"].keys()
+    })
+
+    print("\nRELATÓRIO DE STATUS DO CORREDOR")
+    print("=" * 80)
+
+    total_geral = 0
+    contagem_geral = Counter()
+
+    for grupo, dados in resultado.items():
+        total = dados["total"]
+        contagem = dados["contagem"]
+
+        total_geral += total
+        contagem_geral.update(contagem)
+
+        print(f"\nGrupo: {grupo}")
+        print(f"Total labels: {total}")
+
+        if total == 0:
+            print("  Nenhum label encontrado.")
+            continue
+
+        for status in todos_status:
+            qtd = contagem.get(status, 0)
+            perc = (qtd / total) * 100
+            print(f"  {status:<20} {qtd:>5}  {perc:>6.2f}%")
+
+    print("\nGERAL")
+    print("=" * 80)
+    print(f"Total labels: {total_geral}")
+
+    if total_geral > 0:
+        for status in todos_status:
+            qtd = contagem_geral.get(status, 0)
+            perc = (qtd / total_geral) * 100
+            print(f"  {status:<20} {qtd:>5}  {perc:>6.2f}%")
+
+
+def salvar_csv(resultado, out_csv: Path):
+    import csv
+
+    todos_status = sorted({
+        status
+        for dados in resultado.values()
+        for status in dados["contagem"].keys()
+    })
+
+    with out_csv.open("w", newline="", encoding="utf-8") as f:
+        writer = csv.writer(f)
+        header = ["grupo", "total"]
+
+        for status in todos_status:
+            header.append(f"{status}_qtd")
+            header.append(f"{status}_perc")
+
+        writer.writerow(header)
+
+        for grupo, dados in resultado.items():
+            total = dados["total"]
+            contagem = dados["contagem"]
+
+            row = [grupo, total]
+            for status in todos_status:
+                qtd = contagem.get(status, 0)
+                perc = (qtd / total) * 100 if total > 0 else 0
+                row.extend([qtd, round(perc, 2)])
+
+            writer.writerow(row)
+
+    print(f"\nCSV salvo em: {out_csv}")
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--root",
+        required=True,
+        help="Pasta original/group contendo os grupos com labels/"
+    )
+    parser.add_argument(
+        "--csv",
+        default=None,
+        help="Opcional: caminho para salvar o relatório CSV"
+    )
+
+    args = parser.parse_args()
+
+    root = Path(args.root)
+    resultado = analisar_dataset(root)
+
+    imprimir_relatorio(resultado)
+
+    if args.csv:
+        salvar_csv(resultado, Path(args.csv))
+
+
+if __name__ == "__main__":
+    main()

Python/OAK/datasets/_1_dataset_labeler.py

@@ -0,0 +1,653 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+corridor_state_labeler.py
+-------------------------------------------------
+Rotulador rápido de estado global do corredor para dataset de segmentação.
+
+Objetivo
+- Percorrer uma raiz de dataset no formato:
+
+    dataset_root/
+    ├── grupo_1/
+    │   ├── images/
+    │   ├── masks/
+    │   └── labels/        # criado automaticamente
+    ├── grupo_2/
+    │   ├── images/
+    │   ├── masks/
+    │   └── labels/
+    └── grupo_3/
+        ├── images/
+        ├── masks/
+        └── labels/
+
+- Mostrar imagem original e máscara lado a lado.
+- Permitir classificar cada frame com teclas numéricas.
+- Salvar um JSON por imagem dentro de labels/.
+- Suportar resume, skip, voltar, avançar e undo.
+
+Exemplo de labels globais:
+    Direcionando,EntrandoRua,CaminhandoRua,SaindoRua
+
+Dependências:
+    pip install pillow
+
+Uso com GUI:
+    python corridor_state_labeler.py
+
+Uso por linha de comando:
+    python corridor_state_labeler.py --dataset-root "C:/.../oak-d/dataset/original/group" --states Direcionando EntrandoRua CaminhandoRua SaindoRua --resume
+"""
+
+import argparse
+import json
+import os
+import sys
+from dataclasses import dataclass
+from datetime import datetime
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import tkinter as tk
+from tkinter import filedialog, messagebox
+
+try:
+    from PIL import Image, ImageTk, ImageOps
+except Exception:
+    print("ERRO: Pillow não encontrado. Instale com: pip install pillow", file=sys.stderr)
+    raise
+
+
+IMG_EXTS = [".png", ".jpg", ".jpeg", ".bmp", ".webp"]
+MASK_EXTS = [".png", ".jpg", ".jpeg", ".bmp", ".webp"]
+
+IMAGE_DIR_NAMES = ["images", "image", "imgs", "rgb"]
+MASK_DIR_NAMES = ["masks", "mask", "segmentacao", "seg", "segmentations"]
+LABEL_DIR_NAME = "labels"
+
+
+@dataclass
+class SampleItem:
+    group: str
+    image_path: Path
+    mask_path: Path
+    label_path: Path
+    base: str
+
+
+def normalize_stem(stem: str) -> str:
+    """
+    Remove sufixos comuns para tentar casar image e mask mesmo quando nomes diferem.
+    Exemplos:
+        001_rgb -> 001
+        001_image -> 001
+        001_mask -> 001
+        001_segmentacao -> 001
+    """
+    suffixes = [
+        "_rgb", "_RGB", "_Rgb",
+        "_image", "_img", "_frame",
+        "_mask", "_masks",
+        "_seg", "_SEG", "_segment", "_segmentacao", "_Segmentacao",
+    ]
+    out = stem
+    changed = True
+    while changed:
+        changed = False
+        for sfx in suffixes:
+            if out.endswith(sfx):
+                out = out[: -len(sfx)]
+                changed = True
+                break
+    return out
+
+
+def natural_key(text: str):
+    """Ordenação natural simples: 2 antes de 10."""
+    import re
+    parts = re.split(r"(\d+)", text)
+    return [int(p) if p.isdigit() else p.lower() for p in parts]
+
+
+def find_first_existing_dir(group_dir: Path, candidates: List[str]) -> Optional[Path]:
+    for name in candidates:
+        p = group_dir / name
+        if p.is_dir():
+            return p
+    return None
+
+
+def list_image_files(folder: Path, exts: List[str]) -> List[Path]:
+    files = []
+    for p in folder.iterdir():
+        if p.is_file() and p.suffix.lower() in exts:
+            files.append(p)
+    files.sort(key=lambda x: natural_key(x.name))
+    return files
+
+
+def collect_samples(dataset_root: Path) -> List[SampleItem]:
+    """
+    Procura grupos dentro de dataset_root.
+    Cada grupo precisa ter uma pasta images/ e masks/.
+    Cria labels/ automaticamente.
+    """
+    if not dataset_root.exists() or not dataset_root.is_dir():
+        raise FileNotFoundError(f"Raiz inválida: {dataset_root}")
+
+    samples: List[SampleItem] = []
+
+    group_dirs = [p for p in dataset_root.iterdir() if p.is_dir()]
+    group_dirs.sort(key=lambda x: natural_key(x.name))
+
+    for group_dir in group_dirs:
+        image_dir = find_first_existing_dir(group_dir, IMAGE_DIR_NAMES)
+        mask_dir = find_first_existing_dir(group_dir, MASK_DIR_NAMES)
+
+        if image_dir is None or mask_dir is None:
+            continue
+
+        label_dir = group_dir / LABEL_DIR_NAME
+        label_dir.mkdir(parents=True, exist_ok=True)
+
+        image_files = list_image_files(image_dir, IMG_EXTS)
+        mask_files = list_image_files(mask_dir, MASK_EXTS)
+
+        mask_by_base: Dict[str, Path] = {}
+        for m in mask_files:
+            mask_by_base[normalize_stem(m.stem)] = m
+
+        for img in image_files:
+            base = normalize_stem(img.stem)
+            mask = mask_by_base.get(base)
+            if mask is None:
+                # fallback: tenta mesmo stem exato
+                exact_candidates = [m for m in mask_files if m.stem == img.stem]
+                mask = exact_candidates[0] if exact_candidates else None
+
+            if mask is None:
+                continue
+
+            label_path = label_dir / f"{base}.json"
+            samples.append(
+                SampleItem(
+                    group=group_dir.name,
+                    image_path=img,
+                    mask_path=mask,
+                    label_path=label_path,
+                    base=base,
+                )
+            )
+
+    samples.sort(key=lambda s: (natural_key(s.group), natural_key(s.base)))
+    return samples
+
+
+def read_existing_label(label_path: Path) -> Optional[Dict]:
+    if not label_path.exists():
+        return None
+    try:
+        with label_path.open("r", encoding="utf-8") as f:
+            return json.load(f)
+    except Exception:
+        return None
+
+
+def safe_relative(path: Path, root: Path) -> str:
+    try:
+        return str(path.resolve().relative_to(root.resolve())).replace("\\", "/")
+    except Exception:
+        return str(path).replace("\\", "/")
+
+
+class CorridorStateLabelerApp:
+    def __init__(
+        self,
+        dataset_root: Path,
+        samples: List[SampleItem],
+        states: List[str],
+        resume: bool = True,
+        display_height: int = 520,
+        show_existing: bool = False,
+    ):
+        self.dataset_root = dataset_root
+        self.states = states
+        self.display_height = int(display_height)
+        self.show_existing = show_existing
+
+        if resume:
+            self.samples = [s for s in samples if not s.label_path.exists()]
+        else:
+            self.samples = samples
+
+        self.idx = 0
+        self.history: List[Dict] = []
+
+        self.root = tk.Tk()
+        self.root.title("Agrobot - Corridor State Labeler")
+        self.root.geometry("1600x880")
+        self.root.bind("<Key>", self.on_key)
+
+        self.top_frame = tk.Frame(self.root)
+        self.top_frame.pack(side=tk.TOP, fill=tk.X)
+
+        self.info_label = tk.Label(self.top_frame, text="", font=("Segoe UI", 11), anchor="w")
+        self.info_label.pack(side=tk.LEFT, padx=10, pady=6, fill=tk.X, expand=True)
+
+        self.legend_label = tk.Label(self.top_frame, text=self.build_legend_text(), font=("Segoe UI", 10), anchor="e")
+        self.legend_label.pack(side=tk.RIGHT, padx=10, pady=6)
+
+        self.image_frame = tk.Frame(self.root)
+        self.image_frame.pack(side=tk.TOP, fill=tk.BOTH, expand=True)
+
+        self.left_frame = tk.LabelFrame(self.image_frame, text="Imagem")
+        self.left_frame.pack(side=tk.LEFT, fill=tk.BOTH, expand=True, padx=8, pady=8)
+
+        self.right_frame = tk.LabelFrame(self.image_frame, text="Máscara")
+        self.right_frame.pack(side=tk.LEFT, fill=tk.BOTH, expand=True, padx=8, pady=8)
+
+        self.image_label = tk.Label(self.left_frame)
+        self.image_label.pack(fill=tk.BOTH, expand=True)
+
+        self.mask_label = tk.Label(self.right_frame)
+        self.mask_label.pack(fill=tk.BOTH, expand=True)
+
+        self.current_label_var = tk.StringVar(value="")
+        self.current_label = tk.Label(self.root, textvariable=self.current_label_var, font=("Segoe UI", 11), anchor="w")
+        self.current_label.pack(side=tk.BOTTOM, fill=tk.X, padx=8, pady=2)
+
+        self.status_var = tk.StringVar(value="Pronto.")
+        self.status_label = tk.Label(self.root, textvariable=self.status_var, font=("Segoe UI", 10), anchor="w")
+        self.status_label.pack(side=tk.BOTTOM, fill=tk.X, padx=8, pady=4)
+
+        self.footer = tk.Label(
+            self.root,
+            text="1..9/0=classificar | Espaço=saltar | →/n=próxima | ←/p=anterior | b=desfazer | r=recarregar | q/Esc=sair",
+            font=("Segoe UI", 10),
+        )
+        self.footer.pack(side=tk.BOTTOM, fill=tk.X, pady=2)
+
+        self.render()
+
+    def build_legend_text(self) -> str:
+        parts = []
+        for i, state in enumerate(self.states, start=1):
+            key = i if i <= 9 else 0
+            parts.append(f"[{key}] {state}")
+        return " | ".join(parts)
+
+    def load_and_resize(self, path: Path) -> ImageTk.PhotoImage:
+        img = Image.open(path).convert("RGB")
+        img = ImageOps.exif_transpose(img)
+
+        w, h = img.size
+        if h <= 0:
+            raise ValueError(f"Imagem inválida: {path}")
+
+        new_h = self.display_height
+        new_w = max(1, int(w * (new_h / h)))
+        img = img.resize((new_w, new_h), Image.BILINEAR)
+        return ImageTk.PhotoImage(img)
+
+    def render(self):
+        if not self.samples:
+            messagebox.showinfo("Fim", "Nenhum item para rotular. Talvez tudo já esteja rotulado com resume ligado.")
+            self.root.destroy()
+            return
+
+        self.idx = max(0, min(self.idx, len(self.samples) - 1))
+        item = self.samples[self.idx]
+
+        try:
+            self.tk_img = self.load_and_resize(item.image_path)
+            self.tk_mask = self.load_and_resize(item.mask_path)
+            self.image_label.configure(image=self.tk_img)
+            self.mask_label.configure(image=self.tk_mask)
+        except Exception as e:
+            self.status_var.set(f"ERRO ao abrir imagem/máscara: {e}")
+            return
+
+        existing = read_existing_label(item.label_path)
+        existing_txt = ""
+        if existing:
+            existing_txt = f" | Já rotulado: {existing.get('estado_corredor')}"
+
+        self.info_label.configure(
+            text=(
+                f"{self.idx + 1}/{len(self.samples)} | Grupo: {item.group} | Base: {item.base} "
+                f"| Imagem: {item.image_path.name} | Máscara: {item.mask_path.name}{existing_txt}"
+            )
+        )
+
+        self.current_label_var.set(
+            f"Arquivo de label: {safe_relative(item.label_path, self.dataset_root)}"
+        )
+
+    def save_label(self, state_index: int):
+        if state_index < 0 or state_index >= len(self.states):
+            return
+
+        item = self.samples[self.idx]
+        state = self.states[state_index]
+
+        previous_content = None
+        previous_existed = item.label_path.exists()
+        if previous_existed:
+            try:
+                previous_content = item.label_path.read_text(encoding="utf-8")
+            except Exception:
+                previous_content = None
+
+        rec = {
+            "estado_corredor": state,
+            "label_id": state_index,
+            "states": self.states,
+            "group": item.group,
+            "base": item.base,
+            "image": safe_relative(item.image_path, self.dataset_root),
+            "mask": safe_relative(item.mask_path, self.dataset_root),
+            "label": safe_relative(item.label_path, self.dataset_root),
+            "timestamp": datetime.now().isoformat(timespec="seconds"),
+            "source": "corridor_state_labeler",
+        }
+
+        try:
+            item.label_path.parent.mkdir(parents=True, exist_ok=True)
+            with item.label_path.open("w", encoding="utf-8") as f:
+                json.dump(rec, f, ensure_ascii=False, indent=2)
+
+            self.history.append(
+                {
+                    "action": "label",
+                    "index": self.idx,
+                    "label_path": item.label_path,
+                    "previous_existed": previous_existed,
+                    "previous_content": previous_content,
+                    "new_state": state,
+                }
+            )
+
+            self.status_var.set(f"Salvo: {item.group}/{item.base} → {state}")
+            self.idx += 1
+            if self.idx >= len(self.samples):
+                messagebox.showinfo("Concluído", "Você chegou ao final da fila.")
+                self.root.destroy()
+                return
+            self.render()
+
+        except Exception as e:
+            messagebox.showerror("Erro", f"Falha ao salvar label: {e}")
+            self.status_var.set(f"ERRO ao salvar label: {e}")
+
+    def skip(self):
+        item = self.samples[self.idx]
+        self.history.append({"action": "skip", "index": self.idx})
+        self.status_var.set(f"Pulou: {item.group}/{item.base}")
+        self.idx += 1
+        if self.idx >= len(self.samples):
+            messagebox.showinfo("Concluído", "Você chegou ao final da fila.")
+            self.root.destroy()
+            return
+        self.render()
+
+    def undo(self):
+        if not self.history:
+            self.status_var.set("Nada para desfazer.")
+            return
+
+        last = self.history.pop()
+        action = last.get("action")
+
+        if action == "label":
+            label_path: Path = last["label_path"]
+            try:
+                if last.get("previous_existed"):
+                    previous_content = last.get("previous_content")
+                    if previous_content is not None:
+                        label_path.write_text(previous_content, encoding="utf-8")
+                else:
+                    if label_path.exists():
+                        label_path.unlink()
+
+                self.idx = max(0, min(int(last.get("index", self.idx)), len(self.samples) - 1))
+                self.status_var.set(f"Desfeito label: {label_path.name}")
+                self.render()
+            except Exception as e:
+                messagebox.showerror("Erro", f"Falha ao desfazer: {e}")
+                self.status_var.set(f"ERRO no undo: {e}")
+
+        elif action == "skip":
+            self.idx = max(0, min(int(last.get("index", self.idx)), len(self.samples) - 1))
+            self.status_var.set("Desfeito pulo.")
+            self.render()
+
+    def next_item(self):
+        self.idx = min(len(self.samples) - 1, self.idx + 1)
+        self.status_var.set("Avançou uma imagem.")
+        self.render()
+
+    def prev_item(self):
+        self.idx = max(0, self.idx - 1)
+        self.status_var.set("Voltou uma imagem.")
+        self.render()
+
+    def on_key(self, event):
+        ch = event.keysym.lower()
+
+        if ch in [str(i) for i in range(1, 10)] or ch == "0":
+            idx = 9 if ch == "0" else int(ch) - 1
+            self.save_label(idx)
+            return
+
+        if ch in ("space", "s"):
+            self.skip()
+            return
+
+        if ch in ("right", "n"):
+            self.next_item()
+            return
+
+        if ch in ("left", "p"):
+            self.prev_item()
+            return
+
+        if ch == "b":
+            self.undo()
+            return
+
+        if ch == "r":
+            self.status_var.set("Recarregado.")
+            self.render()
+            return
+
+        if ch in ("q", "escape"):
+            self.root.destroy()
+            return
+
+    def run(self):
+        self.root.mainloop()
+
+
+class SetupWindow:
+    def __init__(self):
+        self.root = tk.Tk()
+        self.root.title("Configurar - Corridor State Labeler")
+        self.root.geometry("760x430")
+
+        frm_root = tk.LabelFrame(self.root, text="Raiz do dataset")
+        frm_root.pack(fill=tk.X, padx=10, pady=8)
+
+        self.dataset_root_var = tk.StringVar(value="")
+        tk.Entry(frm_root, textvariable=self.dataset_root_var).pack(side=tk.LEFT, fill=tk.X, expand=True, padx=6, pady=6)
+        tk.Button(frm_root, text="Escolher...", command=self.choose_dataset_root).pack(side=tk.RIGHT, padx=6, pady=6)
+
+        frm_states = tk.LabelFrame(self.root, text="Estados do corredor separados por vírgula")
+        frm_states.pack(fill=tk.X, padx=10, pady=8)
+
+        self.states_var = tk.StringVar(value="Parado,EntrandoRua,CaminhandoRua,SaindoRua,Manobrando,Direcionando,RetornandoBase,Indefinido")
+        tk.Entry(frm_states, textvariable=self.states_var).pack(fill=tk.X, padx=6, pady=6)
+
+        frm_opts = tk.LabelFrame(self.root, text="Opções")
+        frm_opts.pack(fill=tk.X, padx=10, pady=8)
+
+        self.resume_var = tk.BooleanVar(value=True)
+        self.height_var = tk.IntVar(value=520)
+
+        tk.Checkbutton(frm_opts, text="Retomar: pular imagens que já possuem labels/*.json", variable=self.resume_var).pack(anchor="w", padx=6, pady=4)
+
+        frm_height = tk.Frame(frm_opts)
+        frm_height.pack(fill=tk.X, padx=6, pady=4)
+        tk.Label(frm_height, text="Altura de exibição em px:").pack(side=tk.LEFT)
+        tk.Entry(frm_height, textvariable=self.height_var, width=8).pack(side=tk.LEFT, padx=6)
+
+        self.preview_var = tk.StringVar(value="")
+        tk.Label(self.root, textvariable=self.preview_var, fg="#555", anchor="w", justify="left").pack(fill=tk.X, padx=12, pady=4)
+
+        btn_frame = tk.Frame(self.root)
+        btn_frame.pack(fill=tk.X, padx=10, pady=10)
+        tk.Button(btn_frame, text="Verificar dataset", command=self.preview_dataset).pack(side=tk.LEFT, padx=4)
+        tk.Button(btn_frame, text="Iniciar classificação", command=self.start).pack(side=tk.RIGHT, padx=4)
+
+        tk.Label(
+            self.root,
+            text="Teclas: 1..9/0=classificar | Espaço=saltar | b=desfazer | ←/→ navegar | q/Esc=sair",
+            fg="#555",
+        ).pack(pady=4)
+
+        self.result = None
+
+    def choose_dataset_root(self):
+        p = filedialog.askdirectory(title="Selecione a raiz do dataset")
+        if p:
+            self.dataset_root_var.set(p)
+            self.preview_dataset()
+
+    def parse_states(self) -> List[str]:
+        return [s.strip() for s in self.states_var.get().split(",") if s.strip()]
+
+    def preview_dataset(self):
+        root_raw = self.dataset_root_var.get().strip()
+        if not root_raw:
+            self.preview_var.set("Selecione uma raiz de dataset.")
+            return
+        root = Path(root_raw)
+        try:
+            samples = collect_samples(root)
+            groups = sorted(set(s.group for s in samples), key=natural_key)
+            labeled = sum(1 for s in samples if s.label_path.exists())
+            self.preview_var.set(
+                f"Amostras encontradas: {len(samples)} | Já rotuladas: {labeled} | Grupos: {', '.join(groups) if groups else '-'}"
+            )
+        except Exception as e:
+            self.preview_var.set(f"Erro ao verificar dataset: {e}")
+
+    def start(self):
+        root_raw = self.dataset_root_var.get().strip()
+        states = self.parse_states()
+
+        if not root_raw:
+            messagebox.showwarning("Faltando raiz", "Selecione a raiz do dataset.")
+            return
+        if not states:
+            messagebox.showwarning("Faltando estados", "Informe pelo menos um estado.")
+            return
+        if len(states) > 10:
+            messagebox.showwarning("Muitos estados", "Este rotulador suporta até 10 estados nas teclas 1..9 e 0.")
+            return
+
+        root = Path(root_raw)
+        try:
+            samples = collect_samples(root)
+        except Exception as e:
+            messagebox.showerror("Erro", f"Falha ao ler dataset: {e}")
+            return
+
+        if not samples:
+            messagebox.showinfo(
+                "Sem amostras",
+                "Nenhum par image/mask encontrado. Verifique se cada grupo possui images/ e masks/.",
+            )
+            return
+
+        self.result = {
+            "dataset_root": root,
+            "states": states,
+            "resume": self.resume_var.get(),
+            "display_height": self.height_var.get(),
+        }
+        self.root.destroy()
+
+    def run(self):
+        self.root.mainloop()
+        return self.result
+
+
+def run_gui_setup():
+    setup = SetupWindow()
+    res = setup.run()
+    if not res:
+        return
+
+    dataset_root = res["dataset_root"]
+    samples = collect_samples(dataset_root)
+
+    app = CorridorStateLabelerApp(
+        dataset_root=dataset_root,
+        samples=samples,
+        states=res["states"],
+        resume=res["resume"],
+        display_height=res["display_height"],
+    )
+    app.run()
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Rotulador de estado global do corredor para dataset de segmentação.")
+    parser.add_argument("--dataset-root", help="Raiz do dataset contendo grupos com images/ e masks/.")
+    parser.add_argument("--states", nargs="+", help="Estados do corredor. Ex: Direcionando EntrandoRua CaminhandoRua SaindoRua")
+    parser.add_argument("--resume", action="store_true", help="Pular itens que já possuem labels/*.json.")
+    parser.add_argument("--no-resume", action="store_true", help="Não pular itens já rotulados.")
+    parser.add_argument("--display-height", type=int, default=520, help="Altura de exibição das imagens em px.")
+    parser.add_argument("--no-gui-setup", action="store_true", help="Não abrir janela de configuração.")
+
+    args = parser.parse_args()
+
+    need_gui = not args.no_gui_setup and (not args.dataset_root or not args.states)
+    if need_gui:
+        run_gui_setup()
+        return
+
+    if not args.dataset_root or not args.states:
+        print("ERRO: informe --dataset-root e --states, ou rode sem argumentos para abrir a GUI.", file=sys.stderr)
+        sys.exit(1)
+
+    dataset_root = Path(args.dataset_root)
+    states = args.states
+    if len(states) > 10:
+        print("ERRO: máximo de 10 estados: teclas 1..9 e 0.", file=sys.stderr)
+        sys.exit(1)
+
+    samples = collect_samples(dataset_root)
+    if not samples:
+        print("ERRO: nenhum par image/mask encontrado.", file=sys.stderr)
+        sys.exit(1)
+
+    resume = True
+    if args.no_resume:
+        resume = False
+    elif args.resume:
+        resume = True
+
+    app = CorridorStateLabelerApp(
+        dataset_root=dataset_root,
+        samples=samples,
+        states=states,
+        resume=resume,
+        display_height=args.display_height,
+    )
+    app.run()
+
+
+if __name__ == "__main__":
+    main()

Python/OAK/datasets/_2_create_full_mask.py

@@ -7,7 +7,7 @@ import argparse
 import numpy as np
 
 # ⚙️ Configurações
-with open("config_oak.json", "r") as f:
+with open("config.json", "r") as f:
     config = json.load(f)
 MODELO = config["camera"]
 

Python/OAK/datasets/_4_group_images_by_class.py

@@ -1,22 +1,34 @@
 #!/usr/bin/env python3
 # -*- coding: utf-8 -*-
 """
-Agrupa imagens e máscaras novas em subpastas por combinação de classes presentes.
+Agrupa imagens, máscaras e labels globais em subpastas por combinação de classes presentes.
 
-Estrutura lida (via config.json -> MODELO):
-  MODELO/dataset/original/new_images/
-  MODELO/dataset/original/new_masks/
+Estrutura lida via config.json -> camera:
+
+  MODELO/dataset/original/images/
+  MODELO/dataset/original/masks/
+  MODELO/dataset/original/masks2/     opcional, se dual_head_mask=true
+  MODELO/dataset/original/labels/     opcional/necessário, se dual_head_label=true
 
 Saída:
-  MODELO/dataset/original/group/<grupo>/images
-  MODELO/dataset/original/group/<grupo>/masks
 
-Onde <grupo> é os nomes das classes presentes unidos por "_", ex:
-  chao, erva, cana, chao_erva, erva_cana, chao_erva_cana, etc.
+  MODELO/dataset/original/group/<grupo>/images/
+  MODELO/dataset/original/group/<grupo>/masks/
+  MODELO/dataset/original/group/<grupo>/masks2/    se dual_head_mask=true
+  MODELO/dataset/original/group/<grupo>/labels/    se dual_head_label=true
 
-Requer: utils.carregar_labelmap_completo(labelmap_path)
-O labelmap define mapeamento de cores/ids/nomes das classes.
+Onde <grupo> é o nome das classes presentes na máscara unidos por "_", ex:
+  navegavel
+  naonavegavel
+  naonavegavel_navegavel
+
+Observação importante:
+- dual_head_mask  = segunda cabeça também é máscara pixel-a-pixel.
+- dual_head_label = segunda cabeça é label global por frame, vindo de JSON.
+
+O JSON de label é copiado para labels/ com o mesmo base name final da imagem/máscara.
 """
+
 import os
 import cv2
 import csv
@@ -27,42 +39,60 @@ import numpy as np
 
 from utils import carregar_labelmap_completo, converter_mask_rgb_para_ids
 
+
 # ====================== Configurações base ======================
 
-def carregar_config_e_paths():
-    with open("config_oak.json", "r", encoding="utf-8") as f:
+EXT_IMAGENS = (".jpg", ".jpeg", ".png")
+EXT_MASKS   = (".png", ".jpg", ".jpeg")
+EXT_MASKS2  = (".png", ".jpg", ".jpeg")
+EXT_LABELS  = (".json", ".txt")
+
+MANIFESTO_DEFAULT = "manifest.csv"
+
+
+def carregar_config_e_paths(modelo_cli=None):
+    with open("config.json", "r", encoding="utf-8") as f:
         config = json.load(f)
-    MODELO = config.get("camera")
-    USE_MASKS2 = config.get("dual_head", False)
+
+    MODELO = modelo_cli or config.get("camera")
+    USE_MASKS2 = bool(config.get("dual_head_mask", False))
+    USE_LABELS = bool(config.get("dual_head_label", False))
+
     pasta_base = os.path.join(MODELO, "dataset")
     labelmap_path = os.path.join(pasta_base, "labelmap.txt")
 
-    # Pastas origem/destino
     PASTA_NEW_IMAGES = os.path.join(pasta_base, "original", "images")
     PASTA_NEW_MASKS  = os.path.join(pasta_base, "original", "masks")
     PASTA_NEW_MASKS2 = os.path.join(pasta_base, "original", "masks2")
+    PASTA_NEW_LABELS = os.path.join(pasta_base, "original", "labels")
     PASTA_FINAL      = os.path.join(pasta_base, "original", "group")
 
-    return MODELO, pasta_base, labelmap_path, PASTA_NEW_IMAGES, PASTA_NEW_MASKS, PASTA_NEW_MASKS2, USE_MASKS2, PASTA_FINAL
+    return {
+        "MODELO": MODELO,
+        "pasta_base": pasta_base,
+        "labelmap_path": labelmap_path,
+        "PASTA_NEW_IMAGES": PASTA_NEW_IMAGES,
+        "PASTA_NEW_MASKS": PASTA_NEW_MASKS,
+        "PASTA_NEW_MASKS2": PASTA_NEW_MASKS2,
+        "PASTA_NEW_LABELS": PASTA_NEW_LABELS,
+        "PASTA_FINAL": PASTA_FINAL,
+        "USE_MASKS2": USE_MASKS2,
+        "USE_LABELS": USE_LABELS,
+    }
 
-# Extensões aceitas
-EXT_IMAGENS = (".jpg", ".jpeg", ".png")
-EXT_MASKS   = (".png", ".jpg", ".jpeg")  # prioridade será .png quando houver
-EXT_MASKS2  = (".png", ".jpg", ".jpeg")  # idem
-
-# Manifesto padrão
-MANIFESTO_DEFAULT = "manifest.csv"
 
 # ====================== Utilitários ======================
 
+
 def garantir_pasta(p):
     os.makedirs(p, exist_ok=True)
 
+
 def nome_disponivel(dest_dir, base_name, ext):
-    """Gera nome único em dest_dir com sufixos _001, _002, ... se necessário."""
     cand = os.path.join(dest_dir, base_name + ext)
     if not os.path.exists(cand):
         return cand
+
     i = 1
     while True:
         cand = os.path.join(dest_dir, f"{base_name}_{i:03d}{ext}")
@@ -70,41 +100,69 @@ def nome_disponivel(dest_dir, base_name, ext):
             return cand
         i += 1
 
-def mapear_masks_por_base(pasta_masks):
-    """Retorna {base: caminho_mask}, priorizando .png se houver múltiplas por base."""
+
+def normalizar_base(stem):
+    """
+    Tenta remover sufixos comuns para casar image/mask/label.
+    Ex:
+      001_rgb          -> 001
+      001_segmentacao  -> 001
+      001_mask         -> 001
+    """
+    sufixos = [
+        "_rgb", "_RGB", "_Rgb",
+        "_image", "_img", "_frame",
+        "_mask", "_masks",
+        "_seg", "_SEG", "_segment", "_segmentacao", "_Segmentacao",
+        "_label", "_labels",
+    ]
+
+    out = stem
+    mudou = True
+    while mudou:
+        mudou = False
+        for sfx in sufixos:
+            if out.endswith(sfx):
+                out = out[: -len(sfx)]
+                mudou = True
+                break
+    return out
+
+
+def mapear_arquivos_por_base(pasta, extensoes, normalizar=True):
+    """
+    Retorna {base: caminho}, priorizando .png para máscaras/imagens e .json para labels.
+    """
+    if not pasta or not os.path.isdir(pasta):
+        return {}
+
     mapa = {}
-    for nome in os.listdir(pasta_masks):
+    prioridade = {
+        ".json": 0,
+        ".png": 1,
+        ".jpg": 2,
+        ".jpeg": 3,
+        ".txt": 4,
+    }
+
+    for nome in os.listdir(pasta):
         lower = nome.lower()
-        if not lower.endswith(EXT_MASKS):
+        if not lower.endswith(extensoes):
             continue
-        base, ext = os.path.splitext(nome)
-        cam = os.path.join(pasta_masks, nome)
+
+        stem, ext = os.path.splitext(nome)
+        base = normalizar_base(stem) if normalizar else stem
+        caminho = os.path.join(pasta, nome)
+
         if base not in mapa:
-            mapa[base] = cam
+            mapa[base] = caminho
         else:
             atual_ext = os.path.splitext(mapa[base])[1].lower()
-            if atual_ext != ".png" and ext.lower() == ".png":
-                mapa[base] = cam
+            if prioridade.get(ext.lower(), 99) < prioridade.get(atual_ext, 99):
+                mapa[base] = caminho
+
     return mapa
 
-def mapear_masks2_por_base(pasta_masks2):
-    """Retorna {base: caminho_mask2}, priorizando .png se houver múltiplas por base."""
-    if not pasta_masks2 or not os.path.isdir(pasta_masks2):
-        return {}
-    mapa = {}
-    for nome in os.listdir(pasta_masks2):
-        lower = nome.lower()
-        if not lower.endswith(EXT_MASKS2):
-            continue
-        base, ext = os.path.splitext(nome)
-        cam = os.path.join(pasta_masks2, nome)
-        if base not in mapa:
-            mapa[base] = cam
-        else:
-            atual_ext = os.path.splitext(mapa[base])[1].lower()
-            if atual_ext != ".png" and ext.lower() == ".png":
-                mapa[base] = cam
-    return mapa
 
 def localizar_imagem_por_base(pasta_imgs, base):
     """Retorna caminho da imagem correspondente ao base se existir."""
@@ -112,119 +170,123 @@ def localizar_imagem_por_base(pasta_imgs, base):
         p = os.path.join(pasta_imgs, base + ext)
         if os.path.isfile(p):
             return p
-    return None
+
+    # fallback com normalização
+    mapa = mapear_arquivos_por_base(pasta_imgs, EXT_IMAGENS, normalizar=True)
+    return mapa.get(base)
+
 
 def inferir_ignore_id(ignore_rgb, cor_para_id):
-    """
-    Tenta inferir o ID da classe ignorada a partir do retorno ignore_rgb e do mapa cor->id.
-    - Se ignore_rgb for [id] ou (id,), retorna esse id.
-    - Se ignore_rgb parecer uma cor RGB (len==3), usa cor_para_id[(R,G,B)] se existir.
-    - Caso não consiga, retorna None.
-    """
     if ignore_rgb is None:
         return None
+
     try:
-        # caso [id]
         if isinstance(ignore_rgb, (list, tuple)) and len(ignore_rgb) == 1:
             return int(ignore_rgb[0])
-        # caso [R,G,B]
+
         if isinstance(ignore_rgb, (list, tuple)) and len(ignore_rgb) == 3:
             key = tuple(int(v) for v in ignore_rgb)
             return cor_para_id.get(key)
     except Exception:
         pass
-    # pode já ser um inteiro simples
+
     if isinstance(ignore_rgb, (int, np.integer)):
         return int(ignore_rgb)
+
     return None
 
+
 def extrair_ids_presentes(mask_path, cor_para_id, assume_rgb=True):
-    """
-    Versão corrigida e otimizada:
-
-      - Se máscara for 1 canal: np.unique direto -> IDs.
-      - Se for 3 canais:
-          * se assume_rgb=True: labelmap está em RGB,
-            mas OpenCV lê BGR -> convertemos BGR -> RGB.
-          * se assume_rgb=False: labelmap está em BGR,
-            mantemos BGR como está.
-      - Usa converter_mask_rgb_para_ids em amostragem + fallback full-scan.
-
-    Retorna: set(ids_presentes)
-    """
     m = cv2.imread(mask_path, cv2.IMREAD_UNCHANGED)
     if m is None:
         raise RuntimeError(f"Falha ao abrir máscara: {mask_path}")
 
-    # ---------------------------
-    # CASO 1: máscara indexada
-    # ---------------------------
+    # Máscara indexada
     if len(m.shape) == 2 or (len(m.shape) == 3 and m.shape[2] == 1):
         vals = np.unique(m)
         return set(int(v) for v in vals)
 
-    # ---------------------------
-    # CASO 2: máscara RGB
-    # ---------------------------
-    # Se o labelmap está em RGB (assume_rgb=True),
-    # convertemos a imagem BGR->RGB para casar com as chaves.
+    # Máscara RGB/BGR
     if assume_rgb:
         img = cv2.cvtColor(m, cv2.COLOR_BGR2RGB)
     else:
-        # labelmap já está em BGR; OpenCV entrega BGR; deixa como está
         img = m
 
-    # Aqui as chaves de cor_para_id estão no MESMO espaço de cor da imagem.
-    mapa_rgb = cor_para_id
     max_classes = len(cor_para_id)
 
-    # ---------- AMOSTRAGEM RÁPIDA ----------
-    step = 8  # pode ajustar para 4 se quiser mais precisão
+    # Amostragem rápida
+    step = 8
     amostra = img[::step, ::step]
-    amostra_ids = converter_mask_rgb_para_ids(amostra, mapa_rgb, ignore_id=255)
+    amostra_ids = converter_mask_rgb_para_ids(amostra, cor_para_id, ignore_id=255)
     ids = set(int(x) for x in np.unique(amostra_ids) if x != 255)
 
     if len(ids) >= max_classes:
         return ids
 
-    # ---------- FULL-SCAN (fallback) ----------
-    full_ids = converter_mask_rgb_para_ids(img, mapa_rgb, ignore_id=255)
+    # Full scan
+    full_ids = converter_mask_rgb_para_ids(img, cor_para_id, ignore_id=255)
     ids = set(int(x) for x in np.unique(full_ids) if x != 255)
     return ids
 
+
 def montar_nome_grupo(ids_presentes, id_para_nome):
-    """
-    Constrói o nome do grupo respeitando a ordem natural dos IDs do labelmap.
-    Ex: {0,1,2} -> chao_cana_obstaculo
-    """
-    #print(f"ids_presentes: {ids_presentes}")
     if not ids_presentes:
         return "sem_classe"
+
     nomes = [id_para_nome.get(cid, str(cid)) for cid in sorted(ids_presentes)]
     return "_".join(nomes)
 
-def copiar_ou_mover(img_src, mask_src, dest_img_dir, dest_mask_dir, mover=False, mask2_src=None, dest_mask2_dir=None):
+
+def copiar_arquivo(src, dst, mover=False):
+    if src is None or dst is None:
+        return None
+
+    garantir_pasta(os.path.dirname(dst))
+    if mover:
+        shutil.move(src, dst)
+    else:
+        shutil.copy2(src, dst)
+    return dst
+
+
+def copiar_ou_mover_pacote(
+    img_src,
+    mask_src,
+    dest_img_dir,
+    dest_mask_dir,
+    mover=False,
+    mask2_src=None,
+    dest_mask2_dir=None,
+    label_src=None,
+    dest_label_dir=None,
+):
+    """
+    Copia/move image, mask, mask2 e label usando o MESMO base name final.
+    Isso é essencial para o loader encontrar os pares depois.
+    """
     garantir_pasta(dest_img_dir)
     garantir_pasta(dest_mask_dir)
 
     base_img = os.path.splitext(os.path.basename(img_src))[0]
-    img_ext  = os.path.splitext(img_src)[1].lower()
+    img_ext = os.path.splitext(img_src)[1].lower()
     mask_ext = os.path.splitext(mask_src)[1].lower()
 
     dst_img = nome_disponivel(dest_img_dir, base_img, img_ext)
     new_base = os.path.splitext(os.path.basename(dst_img))[0]
-    dst_mask = os.path.join(dest_mask_dir, new_base + mask_ext)
-    dst_mask2 = None
 
+    dst_mask = os.path.join(dest_mask_dir, new_base + mask_ext)
+
+    # Em caso raro de colisão na mask, recalcula usando ela como referência
     if os.path.exists(dst_mask):
-        # evita colisão invertendo a ordem do "único" para a máscara
         dst_mask = nome_disponivel(dest_mask_dir, new_base, mask_ext)
         new_base = os.path.splitext(os.path.basename(dst_mask))[0]
-        dst_img  = os.path.join(dest_img_dir, new_base + img_ext)
+        dst_img = os.path.join(dest_img_dir, new_base + img_ext)
         if os.path.exists(dst_img):
             dst_img = nome_disponivel(dest_img_dir, new_base, img_ext)
+            new_base = os.path.splitext(os.path.basename(dst_img))[0]
+            dst_mask = os.path.join(dest_mask_dir, new_base + mask_ext)
 
-    # se tiver mask2, usa o MESMO new_base final
+    dst_mask2 = None
     if mask2_src and dest_mask2_dir:
         garantir_pasta(dest_mask2_dir)
         mask2_ext = os.path.splitext(mask2_src)[1].lower()
@@ -232,50 +294,133 @@ def copiar_ou_mover(img_src, mask_src, dest_img_dir, dest_mask_dir, mover=False,
         if os.path.exists(dst_mask2):
             dst_mask2 = nome_disponivel(dest_mask2_dir, new_base, mask2_ext)
 
-    if mover:
-        shutil.move(img_src, dst_img)
-        shutil.move(mask_src, dst_mask)
-        if mask2_src and dst_mask2:
-            shutil.move(mask2_src, dst_mask2)
-    else:
-        shutil.copy2(img_src, dst_img)
-        shutil.copy2(mask_src, dst_mask)
-        if mask2_src and dst_mask2:
-            shutil.copy2(mask2_src, dst_mask2)
+    dst_label = None
+    if label_src and dest_label_dir:
+        garantir_pasta(dest_label_dir)
+        label_ext = os.path.splitext(label_src)[1].lower()
+        dst_label = os.path.join(dest_label_dir, new_base + label_ext)
+        if os.path.exists(dst_label):
+            dst_label = nome_disponivel(dest_label_dir, new_base, label_ext)
+
+    copiar_arquivo(img_src, dst_img, mover=mover)
+    copiar_arquivo(mask_src, dst_mask, mover=mover)
+    copiar_arquivo(mask2_src, dst_mask2, mover=mover)
+    copiar_arquivo(label_src, dst_label, mover=mover)
+
+    return dst_img, dst_mask, dst_mask2, dst_label
+
+
+def validar_dimensoes(img_path, mask_path, mask2_path=None, estrito=False):
+    img = cv2.imread(img_path, cv2.IMREAD_UNCHANGED)
+    msk = cv2.imread(mask_path, cv2.IMREAD_UNCHANGED)
+
+    if img is None or msk is None:
+        raise RuntimeError("Falha ao abrir img/mask.")
+
+    hi, wi = img.shape[:2]
+    hm, wm = msk.shape[:2]
+
+    dim_mismatch = False
+
+    if (hi, wi) != (hm, wm):
+        dim_mismatch = True
+        msg = f"[AVISO] Dimensões diferem: img {wi}x{hi} vs mask {wm}x{hm} para '{os.path.basename(img_path)}'"
+        if estrito:
+            print(msg + " → pulando.")
+            return False, dim_mismatch
+        print(msg + " → copiando mesmo assim.")
+
+    if mask2_path:
+        m2 = cv2.imread(mask2_path, cv2.IMREAD_UNCHANGED)
+        if m2 is None:
+            print(f"[AVISO] Falha ao abrir mask2: {mask2_path} → ignorando mask2.")
+            return True, dim_mismatch
+
+        h2, w2 = m2.shape[:2]
+        if (hi, wi) != (h2, w2):
+            dim_mismatch = True
+            msg2 = f"[AVISO] Dimensões diferem: img {wi}x{hi} vs mask2 {w2}x{h2} para '{os.path.basename(img_path)}'"
+            if estrito:
+                print(msg2 + " → pulando.")
+                return False, dim_mismatch
+            print(msg2 + " → copiando mesmo assim.")
+
+    return True, dim_mismatch
 
-    return dst_img, dst_mask, dst_mask2
 
 # ====================== Pipeline principal ======================
 
-def processar(modelo_cli=None, mover=False, manifesto=MANIFESTO_DEFAULT,
-              validar_dim=True, estrito=False, labelmap_bgr=False):
-    # carrega config/paths
-    MODELO, pasta_base, labelmap_path, PASTA_NEW_IMAGES, PASTA_NEW_MASKS, PASTA_NEW_MASKS2, USE_MASKS2, PASTA_FINAL = carregar_config_e_paths()
 
-    # carrega labelmap completo
+def processar(
+    modelo_cli=None,
+    mover=False,
+    manifesto=MANIFESTO_DEFAULT,
+    validar_dim=True,
+    estrito=False,
+    labelmap_bgr=False,
+    strict_label=False,
+):
+    paths = carregar_config_e_paths(modelo_cli=modelo_cli)
+
+    MODELO = paths["MODELO"]
+    labelmap_path = paths["labelmap_path"]
+    PASTA_NEW_IMAGES = paths["PASTA_NEW_IMAGES"]
+    PASTA_NEW_MASKS = paths["PASTA_NEW_MASKS"]
+    PASTA_NEW_MASKS2 = paths["PASTA_NEW_MASKS2"]
+    PASTA_NEW_LABELS = paths["PASTA_NEW_LABELS"]
+    PASTA_FINAL = paths["PASTA_FINAL"]
+    USE_MASKS2 = paths["USE_MASKS2"]
+    USE_LABELS = paths["USE_LABELS"]
+
+    print(f"[INFO] MODELO: {MODELO}")
+    print(f"[INFO] dual_head_mask: {USE_MASKS2}")
+    print(f"[INFO] dual_head_label: {USE_LABELS}")
+
     cor_para_id, _colormap_rgb, id_para_nome, ignore_rgb = carregar_labelmap_completo(labelmap_path)
     ignore_id = inferir_ignore_id(ignore_rgb, cor_para_id)
-    #print(f"cor_para_id: {cor_para_id}, _colormap_rgb: {_colormap_rgb}, id_para_nome: {id_para_nome}")
 
-    # garante pastas
     garantir_pasta(PASTA_NEW_IMAGES)
     garantir_pasta(PASTA_NEW_MASKS)
     garantir_pasta(PASTA_FINAL)
-    usar_masks2 = USE_MASKS2 and os.path.isdir(PASTA_NEW_MASKS2)
-    if usar_masks2:
-        garantir_pasta(PASTA_NEW_MASKS2)
-        print(f"[INFO] masks2 detectada: {PASTA_NEW_MASKS2}")
 
-    # indexa máscaras
-    mapa_masks = mapear_masks_por_base(PASTA_NEW_MASKS)
-    mapa_masks2 = mapear_masks2_por_base(PASTA_NEW_MASKS2) if usar_masks2 else {}
+    usar_masks2 = USE_MASKS2 and os.path.isdir(PASTA_NEW_MASKS2)
+    usar_labels = USE_LABELS and os.path.isdir(PASTA_NEW_LABELS)
+
+    if USE_MASKS2 and not usar_masks2:
+        print(f"[AVISO] dual_head_mask=true, mas pasta masks2 não existe: {PASTA_NEW_MASKS2}")
+
+    if USE_LABELS and not usar_labels:
+        msg = f"[AVISO] dual_head_label=true, mas pasta labels não existe: {PASTA_NEW_LABELS}"
+        if strict_label:
+            raise RuntimeError(msg)
+        print(msg)
+
+    if usar_masks2:
+        print(f"[INFO] Usando masks2: {PASTA_NEW_MASKS2}")
+
+    if usar_labels:
+        print(f"[INFO] Usando labels: {PASTA_NEW_LABELS}")
+
+    mapa_masks = mapear_arquivos_por_base(PASTA_NEW_MASKS, EXT_MASKS, normalizar=True)
+    mapa_masks2 = mapear_arquivos_por_base(PASTA_NEW_MASKS2, EXT_MASKS2, normalizar=True) if usar_masks2 else {}
+    mapa_labels = mapear_arquivos_por_base(PASTA_NEW_LABELS, EXT_LABELS, normalizar=True) if usar_labels else {}
 
     registros = []
-    totais = {"total_masks":0, "processados":0, "pulados":0, "sem_imagem":0, "sem_mask2":0, "dim_mismatch":0, "erros":0}
+    totais = {
+        "total_masks": 0,
+        "processados": 0,
+        "pulados": 0,
+        "sem_imagem": 0,
+        "sem_mask2": 0,
+        "sem_label": 0,
+        "dim_mismatch": 0,
+        "erros": 0,
+    }
     por_grupo = {}
 
-    for base, mask_path in sorted(mapa_masks.items()):
+    for base, mask_path in sorted(mapa_masks.items(), key=lambda x: x[0]):
         totais["total_masks"] += 1
+
         img_path = localizar_imagem_por_base(PASTA_NEW_IMAGES, base)
         if not img_path:
             totais["sem_imagem"] += 1
@@ -284,114 +429,131 @@ def processar(modelo_cli=None, mover=False, manifesto=MANIFESTO_DEFAULT,
 
         try:
             mask2_path = mapa_masks2.get(base) if usar_masks2 else None
+            label_path = mapa_labels.get(base) if usar_labels else None
+
             if usar_masks2 and not mask2_path:
                 totais["sem_mask2"] += 1
-                print(f"[AVISO] masks2 existe, mas não achei mask2 para base '{base}' (vou agrupar mesmo).")
+                print(f"[AVISO] masks2 ativada, mas não achei mask2 para base '{base}'. Vou agrupar sem mask2.")
+
+            if usar_labels and not label_path:
+                totais["sem_label"] += 1
+                msg = f"[AVISO] labels ativada, mas não achei label para base '{base}'."
+                if strict_label:
+                    print(msg + " → pulando.")
+                    totais["pulados"] += 1
+                    continue
+                print(msg + " Vou agrupar sem label.")
 
             if validar_dim:
-                try:
-                    img = cv2.imread(img_path, cv2.IMREAD_UNCHANGED)
-                    msk = cv2.imread(mask_path, cv2.IMREAD_UNCHANGED)
-                    if img is None or msk is None:
-                        raise RuntimeError("Falha ao abrir img/mask.")
-                    hi, wi = img.shape[:2]
-                    hm, wm = msk.shape[:2]
-                    if (hi, wi) != (hm, wm):
-                        totais["dim_mismatch"] += 1
-                        msg = f"[AVISO] Dimensões diferem (img {wi}x{hi} vs mask {wm}x{hm}) para base '{base}'"
-                        if estrito:
-                            print(msg + "  → pulando.")
-                            totais["pulados"] += 1
-                            continue
-                        else:
-                            print(msg + "  → copiando mesmo assim.")
-
-                    if mask2_path:
-                        m2 = cv2.imread(mask2_path, cv2.IMREAD_UNCHANGED)
-                        if m2 is None:
-                            print(f"[AVISO] Falha ao abrir mask2: {mask2_path}  → ignorando mask2.")
-                            mask2_path = None
-                        else:
-                            h2, w2 = m2.shape[:2]
-                            if (hi, wi) != (h2, w2):
-                                totais["dim_mismatch"] += 1
-                                msg2 = f"[AVISO] Dimensões diferem (img {wi}x{hi} vs mask2 {w2}x{h2}) para base '{base}'"
-                                if estrito:
-                                    print(msg2 + "  → pulando (por mask2).")
-                                    totais["pulados"] += 1
-                                    continue
-                                else:
-                                    print(msg2 + "  → copiando mesmo assim.")
-
-                except Exception as e_dim:
-                    print(f"[AVISO] Falha ao validar dimensões: {e_dim}  → copiando mesmo assim.")
+                ok_dim, dim_mismatch = validar_dimensoes(img_path, mask_path, mask2_path=mask2_path, estrito=estrito)
+                if dim_mismatch:
+                    totais["dim_mismatch"] += 1
+                if not ok_dim:
+                    totais["pulados"] += 1
+                    continue
 
             ids_presentes = extrair_ids_presentes(mask_path, cor_para_id, assume_rgb=not labelmap_bgr)
-            # remove classe ignorada, se conhecida
             if ignore_id is not None and ignore_id in ids_presentes:
                 ids_presentes.discard(ignore_id)
 
-            # monta nome do grupo
             grupo = montar_nome_grupo(ids_presentes, id_para_nome)
 
-            # destinos
             dest_base = os.path.join(PASTA_FINAL, grupo)
-            dest_img_dir  = os.path.join(dest_base, "images")
+            dest_img_dir = os.path.join(dest_base, "images")
             dest_mask_dir = os.path.join(dest_base, "masks")
             dest_mask2_dir = os.path.join(dest_base, "masks2") if usar_masks2 else None
+            dest_label_dir = os.path.join(dest_base, "labels") if usar_labels else None
 
-            dst_img, dst_mask, dst_mask2 = copiar_ou_mover(img_path, mask_path, dest_img_dir, dest_mask_dir, mover=mover, mask2_src=mask2_path, dest_mask2_dir=dest_mask2_dir)
+            dst_img, dst_mask, dst_mask2, dst_label = copiar_ou_mover_pacote(
+                img_src=img_path,
+                mask_src=mask_path,
+                dest_img_dir=dest_img_dir,
+                dest_mask_dir=dest_mask_dir,
+                mover=mover,
+                mask2_src=mask2_path,
+                dest_mask2_dir=dest_mask2_dir,
+                label_src=label_path,
+                dest_label_dir=dest_label_dir,
+            )
 
             totais["processados"] += 1
-            registros.append([img_path, mask_path, mask2_path or "", dst_img, dst_mask, dst_mask2 or "", grupo])
             por_grupo[grupo] = por_grupo.get(grupo, 0) + 1
 
-            print(f"[OK] {os.path.basename(dst_img)}  → grupo: {grupo}")
-            extra = " +mask2" if dst_mask2 else ""
-            print(f"[OK] {os.path.basename(dst_img)}{extra}  → grupo: {grupo}")
+            registros.append([
+                img_path,
+                mask_path,
+                mask2_path or "",
+                label_path or "",
+                dst_img,
+                dst_mask,
+                dst_mask2 or "",
+                dst_label or "",
+                grupo,
+            ])
+
+            extras = []
+            if dst_mask2:
+                extras.append("mask2")
+            if dst_label:
+                extras.append("label")
+            extra_txt = " + " + " + ".join(extras) if extras else ""
+            print(f"[OK] {os.path.basename(dst_img)}{extra_txt} → grupo: {grupo}")
 
         except Exception as e:
             totais["erros"] += 1
             print(f"[ERRO] base '{base}': {e}")
 
-    # manifesto
     if manifesto and registros:
         with open(manifesto, "w", newline="", encoding="utf-8") as f:
             w = csv.writer(f)
-            w.writerow(["src_image", "src_mask", "src_mask2", "dst_image", "dst_mask", "dst_mask2", "grupo"])
+            w.writerow([
+                "src_image",
+                "src_mask",
+                "src_mask2",
+                "src_label",
+                "dst_image",
+                "dst_mask",
+                "dst_mask2",
+                "dst_label",
+                "grupo",
+            ])
             w.writerows(registros)
         print(f"[MANIFESTO] {manifesto} salvo ({len(registros)} entradas).")
 
-    # resumo
     print("\nResumo: " + " | ".join(f"{k}={v}" for k, v in totais.items()))
     if por_grupo:
         print("Por grupo:")
         for g, c in sorted(por_grupo.items(), key=lambda x: x[0]):
             print(f"  - {g}: {c}")
 
+
 # ====================== CLI ======================
 
+
 def build_cli():
     ap = argparse.ArgumentParser(
-        description="Agrupa pares IMG+MASK por combinação de classes presentes na máscara (a partir de new_*)."
+        description="Agrupa IMG+MASK e opcionalmente MASK2/LABEL por combinação de classes presentes na máscara."
     )
-    ap.add_argument("--move", action="store_true", help="Move (em vez de copiar) para as pastas de grupo.")
-    ap.add_argument("--manifest", default=MANIFESTO_DEFAULT, help="CSV de manifesto ('' para não gerar).")
-    ap.add_argument("--modelo", default=None, help="Sobrescreve MODELO do config.json.")
+    ap.add_argument("--move", action="store_true", help="Move em vez de copiar para as pastas de grupo.")
+    ap.add_argument("--manifest", default=MANIFESTO_DEFAULT, help="CSV de manifesto. Use '' para não gerar.")
+    ap.add_argument("--modelo", default=None, help="Sobrescreve camera do config.json.")
     ap.add_argument("--no-validate", action="store_true", help="Não validar dimensões de imagem/máscara.")
     ap.add_argument("--strict", action="store_true", help="Se validar e forem diferentes, pular o par.")
-    ap.add_argument("--labels-bgr", action="store_true",
-                    help="Use se o labelmap estiver em BGR (por padrão assume RGB).")
+    ap.add_argument("--labels-bgr", action="store_true", help="Use se o labelmap estiver em BGR. Por padrão assume RGB.")
+    ap.add_argument("--strict-label", action="store_true", help="Se dual_head_label=true e faltar label, pula o item.")
     return ap
 
+
 if __name__ == "__main__":
     args = build_cli().parse_args()
     manifest = None if (args.manifest.strip() == "") else args.manifest
+
     processar(
         modelo_cli=args.modelo,
         mover=args.move,
         manifesto=manifest,
         validar_dim=not args.no_validate,
         estrito=args.strict,
-        labelmap_bgr=args.labels_bgr
+        labelmap_bgr=args.labels_bgr,
+        strict_label=args.strict_label,
     )

Python/OAK/datasets/_5_augmentation.py

@@ -1,113 +1,185 @@
 #!/usr/bin/env python3
 # -*- coding: utf-8 -*-
 """
-Augmenta imagens e máscaras *por grupo*.
+Augmenta imagens, máscaras e opcionalmente labels globais por grupo.
 
-Entrada (via config.json -> MODELO):
-  MODELO/dataset/original/group/<grupo>/images
-  MODELO/dataset/original/group/<grupo>/masks
+Entrada via config_oak.json -> camera:
+
+  MODELO/dataset/original/group/<grupo>/images/
+  MODELO/dataset/original/group/<grupo>/masks/
+  MODELO/dataset/original/group/<grupo>/masks2/    opcional, se dual_head_mask=true
+  MODELO/dataset/original/group/<grupo>/labels/    opcional, se dual_head_label=true
 
 Saída:
-  MODELO/dataset/augmented/group/<grupo>/images
-  MODELO/dataset/augmented/group/<grupo>/masks
 
-Se "original/group" não existir, faz fallback para:
-  MODELO/dataset/original/{images,masks}
-  MODELO/dataset/augmented/{images,masks}
+  MODELO/dataset/augmented/group/<grupo>/images/
+  MODELO/dataset/augmented/group/<grupo>/masks/
+  MODELO/dataset/augmented/group/<grupo>/masks2/    se dual_head_mask=true
+  MODELO/dataset/augmented/group/<grupo>/labels/    se dual_head_label=true
 
-Transf. geométricas (aplicam a img e máscara) e fotométricas (apenas imagem).
+Observação:
+- A máscara e a mask2 recebem as mesmas transformações geométricas da imagem.
+- O label global NÃO é transformado visualmente. Ele é copiado e atualizado para apontar para o novo base augmentado.
 
 Uso:
-  python _3_augmentation_grouped.py --copies 5
-  python _3_augmentation_grouped.py --copies 5 --groups chao,chao_erva,cana
+  python _3_augmentation_grouped_with_labels.py --copies 5
+  python _3_augmentation_grouped_with_labels.py --copies 5 --groups navegavel,naonavegavel_navegavel
+  python _3_augmentation_grouped_with_labels.py --copies 5 --strict-label
 """
+
 import os
 import json
+import argparse
+from pathlib import Path
+
 import cv2
 from PIL import Image
 import albumentations as A
-import argparse
 
-# ⚙️ Configurações
-with open("config_oak.json", "r", encoding="utf-8") as f:
+
+# ====================== Configurações ======================
+
+with open("config.json", "r", encoding="utf-8") as f:
     config = json.load(f)
-MODELO = config.get("camera", ".")
-USE_MASKS2 = config.get("dual_head", False)
 
-# Pastas base
+MODELO = config.get("camera", ".")
+USE_MASKS2 = bool(config.get("dual_head_mask", False))
+USE_LABELS = bool(config.get("dual_head_label", False))
+
 DATASET_BASE = os.path.join(MODELO, "dataset")
 ORIG_GROUP_ROOT = os.path.join(DATASET_BASE, "original", "group")
-AUG_GROUP_ROOT  = os.path.join(DATASET_BASE, "augmented", "group")
+AUG_GROUP_ROOT = os.path.join(DATASET_BASE, "augmented", "group")
 
-# Fallback (modo antigo, sem grupos)
+# Fallback legacy, sem grupos
 ORIG_OLD_IMG = os.path.join(DATASET_BASE, "original", "images")
 ORIG_OLD_MSK = os.path.join(DATASET_BASE, "original", "masks")
 ORIG_OLD_MSK2 = os.path.join(DATASET_BASE, "original", "masks2")
-AUG_OLD_IMG  = os.path.join(DATASET_BASE, "augmented", "images")
-AUG_OLD_MSK  = os.path.join(DATASET_BASE, "augmented", "masks")
-AUG_OLD_MSK2 = os.path.join(DATASET_BASE, "augmented", "masks2")
+ORIG_OLD_LABELS = os.path.join(DATASET_BASE, "original", "labels")
+
+AUG_OLD_IMG = os.path.join(DATASET_BASE, "augmented", "images")
+AUG_OLD_MSK = os.path.join(DATASET_BASE, "augmented", "masks")
+AUG_OLD_MSK2 = os.path.join(DATASET_BASE, "augmented", "masks2")
+AUG_OLD_LABELS = os.path.join(DATASET_BASE, "augmented", "labels")
 
-# Extensões aceitas
 IMG_EXTS = (".jpg", ".jpeg", ".png")
-MSK_EXTS = (".png", ".jpg", ".jpeg")  # manter prioridade PNG quando possível
+MSK_EXTS = (".png", ".jpg", ".jpeg")
 MSK2_EXTS = (".png", ".jpg", ".jpeg")
+LABEL_EXTS = (".json", ".txt")
+
+
+# ====================== Augmentation ======================
+
+train_tf = A.Compose(
+    [
+        A.HorizontalFlip(p=0.5),
+
+        A.ShiftScaleRotate(
+            shift_limit=0.01,
+            scale_limit=0.10,
+            rotate_limit=5,
+            border_mode=cv2.BORDER_REFLECT_101,
+            interpolation=cv2.INTER_LINEAR,
+            p=0.30,
+        ),
+
+        A.OneOf(
+            [
+                A.RandomBrightnessContrast(0.2, 0.2, p=1.0),
+                A.HueSaturationValue(hue_shift_limit=5, sat_shift_limit=20, val_shift_limit=15, p=1.0),
+                A.RandomGamma(gamma_limit=(90, 110), p=1.0),
+            ],
+            p=0.70,
+        ),
+
+        A.OneOf(
+            [
+                A.MotionBlur(blur_limit=3, p=1.0),
+                A.GaussianBlur(blur_limit=3, p=1.0),
+            ],
+            p=0.20,
+        ),
+
+        A.OneOf(
+            [
+                A.GaussNoise(var_limit=(5.0, 15.0), p=1.0),
+                A.ImageCompression(quality_lower=50, quality_upper=85, p=1.0),
+            ],
+            p=0.20,
+        ),
+
+        A.RandomShadow(p=0.10),
+        A.RandomSunFlare(p=0.10),
+        A.ChannelShuffle(p=0.05),
+        A.CoarseDropout(max_holes=6, max_height=16, max_width=16, p=0.10),
+    ],
+    additional_targets={"mask2": "mask"},
+)
+
+
+# ====================== Utilitários ======================
+
 
 def garantir_dir(p):
     os.makedirs(p, exist_ok=True)
 
-# Pipeline de augmentations
-train_tf = A.Compose([
-    A.HorizontalFlip(p=0.5),
 
-    # Geométricas (aplicam em imagem e máscara)
-    A.ShiftScaleRotate(
-        shift_limit=0.01,
-        scale_limit=0.10,
-        rotate_limit=5,
-        border_mode=cv2.BORDER_REFLECT_101,
-        interpolation=cv2.INTER_LINEAR,
-        p=0.30
-    ),
+def normalizar_base(stem: str) -> str:
+    sufixos = [
+        "_rgb", "_RGB", "_Rgb",
+        "_image", "_img", "_frame",
+        "_mask", "_masks",
+        "_seg", "_SEG", "_segment", "_segmentacao", "_Segmentacao",
+        "_label", "_labels",
+    ]
 
-    # Fotométricas (somente imagem)
-    A.OneOf([
-        A.RandomBrightnessContrast(0.2, 0.2, p=1.0),
-        A.HueSaturationValue(hue_shift_limit=5, sat_shift_limit=20, val_shift_limit=15, p=1.0),
-        A.RandomGamma(gamma_limit=(90, 110), p=1.0),
-    ], p=0.70),
+    out = stem
+    mudou = True
+    while mudou:
+        mudou = False
+        for sfx in sufixos:
+            if out.endswith(sfx):
+                out = out[: -len(sfx)]
+                mudou = True
+                break
+    return out
 
-    A.OneOf([
-        A.MotionBlur(blur_limit=3, p=1.0),
-        A.GaussianBlur(blur_limit=3, p=1.0),
-    ], p=0.20),
 
-    A.OneOf([
-        A.GaussNoise(var_limit=(5.0, 15.0), p=1.0),
-        A.ImageCompression(quality_lower=50, quality_upper=85, p=1.0),
-    ], p=0.20),
+def map_files_by_base(folder, exts):
+    by_base = {}
+    if not os.path.isdir(folder):
+        return by_base
 
-    A.RandomShadow(p=0.10),
-    A.RandomSunFlare(p=0.10),
-    A.ChannelShuffle(p=0.05),
-    A.CoarseDropout(max_holes=6, max_height=16, max_width=16, p=0.10),
-], additional_targets={
-    'mask2': 'mask'
-})
+    prioridade = {
+        ".json": 0,
+        ".png": 1,
+        ".jpg": 2,
+        ".jpeg": 3,
+        ".txt": 4,
+    }
 
-def load_rgb(path):
-    # cv2 lê BGR → converte para RGB
-    im = cv2.imread(path, cv2.IMREAD_COLOR)
-    if im is None:
-        raise FileNotFoundError(path)
-    return cv2.cvtColor(im, cv2.COLOR_BGR2RGB)
+    for fname in os.listdir(folder):
+        lower = fname.lower()
+        if not lower.endswith(exts):
+            continue
+
+        stem, ext = os.path.splitext(fname)
+        base = normalizar_base(stem)
+        cand = os.path.join(folder, fname)
+
+        if base not in by_base:
+            by_base[base] = cand
+        else:
+            cur_ext = os.path.splitext(by_base[base])[1].lower()
+            if prioridade.get(ext.lower(), 99) < prioridade.get(cur_ext, 99):
+                by_base[base] = cand
+
+    return by_base
 
-def save_rgb(path, arr_rgb):
-    Image.fromarray(arr_rgb).save(path)
 
 def list_groups(root):
-    """Lista grupos válidos (que contêm subpastas images e masks)."""
     if not os.path.isdir(root):
         return []
+
     grupos = []
     for name in sorted(os.listdir(root)):
         gdir = os.path.join(root, name)
@@ -117,68 +189,107 @@ def list_groups(root):
             grupos.append(name)
     return grupos
 
-def map_masks_by_base(msk_dir):
-    """Mapeia máscaras por base (prioriza .png)."""
-    by_base = {}
-    if not os.path.isdir(msk_dir):
-        return by_base
-    for fname in os.listdir(msk_dir):
-        f_lower = fname.lower()
-        if not f_lower.endswith(MSK_EXTS):
-            continue
-        base, ext = os.path.splitext(fname)
-        cand = os.path.join(msk_dir, fname)
-        if base not in by_base:
-            by_base[base] = cand
-        else:
-            # mantém .png se disponível
-            cur_ext = os.path.splitext(by_base[base])[1].lower()
-            if cur_ext != ".png" and ext.lower() == ".png":
-                by_base[base] = cand
-    return by_base
 
-def map_masks2_by_base(msk2_dir):
-    """Mapeia máscaras2 por base (prioriza .png)."""
-    by_base = {}
-    if not os.path.isdir(msk2_dir):
-        return by_base
-    for fname in os.listdir(msk2_dir):
-        f_lower = fname.lower()
-        if not f_lower.endswith(MSK2_EXTS):
-            continue
-        base, ext = os.path.splitext(fname)
-        cand = os.path.join(msk2_dir, fname)
-        if base not in by_base:
-            by_base[base] = cand
-        else:
-            cur_ext = os.path.splitext(by_base[base])[1].lower()
-            if cur_ext != ".png" and ext.lower() == ".png":
-                by_base[base] = cand
-    return by_base
+def load_rgb(path):
+    im = cv2.imread(path, cv2.IMREAD_COLOR)
+    if im is None:
+        raise FileNotFoundError(path)
+    return cv2.cvtColor(im, cv2.COLOR_BGR2RGB)
 
-def ensure_aug_dirs(group_name=None, use_masks2=False):
-    """Cria diretórios de saída para o grupo ou modo antigo. Se use_masks2, cria masks2."""
+
+def save_rgb(path, arr_rgb):
+    garantir_dir(os.path.dirname(path))
+    Image.fromarray(arr_rgb).save(path)
+
+
+def ensure_aug_dirs(group_name=None, use_masks2=False, use_labels=False):
     if group_name:
         img_out = os.path.join(AUG_GROUP_ROOT, group_name, "images")
         msk_out = os.path.join(AUG_GROUP_ROOT, group_name, "masks")
         msk2_out = os.path.join(AUG_GROUP_ROOT, group_name, "masks2") if use_masks2 else None
+        labels_out = os.path.join(AUG_GROUP_ROOT, group_name, "labels") if use_labels else None
     else:
         img_out = AUG_OLD_IMG
         msk_out = AUG_OLD_MSK
         msk2_out = AUG_OLD_MSK2 if use_masks2 else None
+        labels_out = AUG_OLD_LABELS if use_labels else None
+
     garantir_dir(img_out)
     garantir_dir(msk_out)
+
     if use_masks2 and msk2_out:
         garantir_dir(msk2_out)
-    return img_out, msk_out, msk2_out
 
-def augment_pair(img_path, msk_path, img_out_dir, msk_out_dir, copies, msk2_path=None, msk2_out_dir=None):
+    if use_labels and labels_out:
+        garantir_dir(labels_out)
+
+    return img_out, msk_out, msk2_out, labels_out
+
+
+def safe_rel(path, root):
+    try:
+        return str(Path(path).resolve().relative_to(Path(root).resolve())).replace("\\", "/")
+    except Exception:
+        return str(path).replace("\\", "/")
+
+
+def copiar_label_aug(label_path, out_label_path, new_base, out_img_path, out_msk_path, out_msk2_path=None, group_name=None):
+    """
+    Copia label global para a amostra augmentada.
+    Se for JSON, atualiza campos úteis.
+    Se for TXT, copia o conteúdo como está.
+    """
+    if not label_path or not out_label_path:
+        return
+
+    garantir_dir(os.path.dirname(out_label_path))
+    ext = os.path.splitext(label_path)[1].lower()
+
+    if ext == ".json":
+        try:
+            with open(label_path, "r", encoding="utf-8") as f:
+                data = json.load(f)
+        except Exception:
+            data = {}
+
+        data["base"] = new_base
+        data["group"] = group_name if group_name is not None else data.get("group")
+        data["image"] = safe_rel(out_img_path, DATASET_BASE)
+        data["mask"] = safe_rel(out_msk_path, DATASET_BASE)
+        data["label"] = safe_rel(out_label_path, DATASET_BASE)
+        data["is_augmented"] = True
+        data["source_label"] = safe_rel(label_path, DATASET_BASE)
+
+        if out_msk2_path:
+            data["mask2"] = safe_rel(out_msk2_path, DATASET_BASE)
+
+        with open(out_label_path, "w", encoding="utf-8") as f:
+            json.dump(data, f, ensure_ascii=False, indent=2)
+    else:
+        with open(label_path, "r", encoding="utf-8") as f:
+            content = f.read()
+        with open(out_label_path, "w", encoding="utf-8") as f:
+            f.write(content)
+
+
+def augment_pair(
+    img_path,
+    msk_path,
+    img_out_dir,
+    msk_out_dir,
+    copies,
+    msk2_path=None,
+    msk2_out_dir=None,
+    label_path=None,
+    label_out_dir=None,
+    group_name=None,
+):
     base_img, img_ext = os.path.splitext(os.path.basename(img_path))
-    base_msk, msk_ext = os.path.splitext(os.path.basename(msk_path))
+    _, msk_ext = os.path.splitext(os.path.basename(msk_path))
     msk2_ext = os.path.splitext(os.path.basename(msk2_path))[1] if msk2_path else None
+    label_ext = os.path.splitext(os.path.basename(label_path))[1] if label_path else None
 
-    # padroniza pelo base da imagem
-    base = base_img
+    base = normalizar_base(base_img)
 
     img = load_rgb(img_path)
     msk = load_rgb(msk_path)
@@ -194,122 +305,223 @@ def augment_pair(img_path, msk_path, img_out_dir, msk_out_dir, copies, msk2_path
         img_aug = aug["image"]
         msk_aug = aug["mask"]
 
-        out_img = os.path.join(img_out_dir, f"{base}_aug_{i:02d}{img_ext}")
-        out_msk = os.path.join(msk_out_dir, f"{base}_aug_{i:02d}{msk_ext}")
+        new_base = f"{base}_aug_{i:02d}"
+        out_img = os.path.join(img_out_dir, f"{new_base}{img_ext}")
+        out_msk = os.path.join(msk_out_dir, f"{new_base}{msk_ext}")
+
         save_rgb(out_img, img_aug)
         save_rgb(out_msk, msk_aug)
 
+        out_msk2 = None
         if msk2 is not None and msk2_out_dir:
             msk2_aug = aug["mask2"]
-            out_msk2 = os.path.join(msk2_out_dir, f"{base}_aug_{i:02d}{msk2_ext}")
+            out_msk2 = os.path.join(msk2_out_dir, f"{new_base}{msk2_ext}")
             save_rgb(out_msk2, msk2_aug)
 
+        if label_path and label_out_dir:
+            out_label = os.path.join(label_out_dir, f"{new_base}{label_ext}")
+            copiar_label_aug(
+                label_path=label_path,
+                out_label_path=out_label,
+                new_base=new_base,
+                out_img_path=out_img,
+                out_msk_path=out_msk,
+                out_msk2_path=out_msk2,
+                group_name=group_name,
+            )
+
         gen += 1
+
     return gen
 
-def process_group(group_name, copies):
-    """Processa um grupo único (images/masks dentro de ORIG_GROUP_ROOT/<group_name>/)."""
+
+# ====================== Processamento ======================
+
+
+def process_group(group_name, copies, strict_label=False):
     img_dir = os.path.join(ORIG_GROUP_ROOT, group_name, "images")
     msk_dir = os.path.join(ORIG_GROUP_ROOT, group_name, "masks")
     msk2_dir = os.path.join(ORIG_GROUP_ROOT, group_name, "masks2")
+    labels_dir = os.path.join(ORIG_GROUP_ROOT, group_name, "labels")
+
     if not (os.path.isdir(img_dir) and os.path.isdir(msk_dir)):
-        print(f"[WARN] Grupo '{group_name}' inválido (sem images/masks). Pulando.")
+        print(f"[WARN] Grupo '{group_name}' inválido, sem images/masks. Pulando.")
         return 0
 
     imgs = [f for f in os.listdir(img_dir) if os.path.splitext(f.lower())[1] in IMG_EXTS]
-    msk_map = map_masks_by_base(msk_dir)
+
+    msk_map = map_files_by_base(msk_dir, MSK_EXTS)
+
     use_masks2 = USE_MASKS2 and os.path.isdir(msk2_dir)
-    msk2_map = map_masks2_by_base(msk2_dir) if use_masks2 else {}
-    img_out_dir, msk_out_dir, msk2_out_dir = ensure_aug_dirs(group_name, use_masks2=use_masks2)
+    msk2_map = map_files_by_base(msk2_dir, MSK2_EXTS) if use_masks2 else {}
+
+    use_labels = USE_LABELS and os.path.isdir(labels_dir)
+    label_map = map_files_by_base(labels_dir, LABEL_EXTS) if use_labels else {}
+
+    if USE_MASKS2 and not use_masks2:
+        print(f"[WARN] [{group_name}] dual_head_mask=true, mas pasta masks2 não existe.")
+
+    if USE_LABELS and not use_labels:
+        msg = f"[WARN] [{group_name}] dual_head_label=true, mas pasta labels não existe."
+        if strict_label:
+            print(msg + " Pulando grupo.")
+            return 0
+        print(msg + " Gerando sem labels.")
+
+    img_out_dir, msk_out_dir, msk2_out_dir, label_out_dir = ensure_aug_dirs(
+        group_name,
+        use_masks2=use_masks2,
+        use_labels=use_labels,
+    )
 
     count = 0
+    sem_mask = 0
+    sem_mask2 = 0
+    sem_label = 0
+
     for img_file in sorted(imgs):
         base, _ = os.path.splitext(img_file)
-        msk_file = msk_map.get(base)
+        base_norm = normalizar_base(base)
+
+        msk_file = msk_map.get(base_norm)
         if not msk_file:
+            sem_mask += 1
             print(f"[WARN] [{group_name}] Máscara não encontrada para {img_file}, pulando.")
             continue
-        msk2_file = msk2_map.get(base) if use_masks2 else None
+
+        msk2_file = msk2_map.get(base_norm) if use_masks2 else None
         if use_masks2 and not msk2_file:
+            sem_mask2 += 1
             print(f"[WARN] [{group_name}] mask2 não encontrada para {img_file}, gerando só img+mask.")
+
+        label_file = label_map.get(base_norm) if use_labels else None
+        if use_labels and not label_file:
+            sem_label += 1
+            msg = f"[WARN] [{group_name}] label não encontrado para {img_file}."
+            if strict_label:
+                print(msg + " Pulando item.")
+                continue
+            print(msg + " Gerando augmentation sem label.")
+
         try:
             count += augment_pair(
-                os.path.join(img_dir, img_file),
-                msk_file,
-                img_out_dir,
-                msk_out_dir,
+                img_path=os.path.join(img_dir, img_file),
+                msk_path=msk_file,
+                img_out_dir=img_out_dir,
+                msk_out_dir=msk_out_dir,
                 copies=copies,
                 msk2_path=msk2_file,
-                msk2_out_dir=msk2_out_dir
+                msk2_out_dir=msk2_out_dir,
+                label_path=label_file,
+                label_out_dir=label_out_dir,
+                group_name=group_name,
             )
         except Exception as e:
             print(f"[ERRO] [{group_name}] {img_file}: {e}")
-    print(f"[OK] Grupo '{group_name}' → {count} pares gerados.")
+
+    print(
+        f"[OK] Grupo '{group_name}' → {count} pares gerados. "
+        f"sem_mask={sem_mask} | sem_mask2={sem_mask2} | sem_label={sem_label}"
+    )
     return count
 
-def process_legacy(copies):
-    """Fallback: modo sem grupos (original/images e original/masks)."""
+
+def process_legacy(copies, strict_label=False):
     if not (os.path.isdir(ORIG_OLD_IMG) and os.path.isdir(ORIG_OLD_MSK)):
         print("[WARN] Modo legacy não encontrado. Nada a fazer.")
         return 0
 
     imgs = [f for f in os.listdir(ORIG_OLD_IMG) if os.path.splitext(f.lower())[1] in IMG_EXTS]
-    msk_map = map_masks_by_base(ORIG_OLD_MSK)
+
+    msk_map = map_files_by_base(ORIG_OLD_MSK, MSK_EXTS)
+
     use_masks2 = USE_MASKS2 and os.path.isdir(ORIG_OLD_MSK2)
-    msk2_map = map_masks2_by_base(ORIG_OLD_MSK2) if use_masks2 else {}
-    img_out_dir, msk_out_dir, msk2_out_dir = ensure_aug_dirs(group_name=None, use_masks2=use_masks2)
+    msk2_map = map_files_by_base(ORIG_OLD_MSK2, MSK2_EXTS) if use_masks2 else {}
+
+    use_labels = USE_LABELS and os.path.isdir(ORIG_OLD_LABELS)
+    label_map = map_files_by_base(ORIG_OLD_LABELS, LABEL_EXTS) if use_labels else {}
+
+    if USE_LABELS and not use_labels and strict_label:
+        print("[WARN] Legacy com dual_head_label=true, mas sem pasta labels. Pulando.")
+        return 0
+
+    img_out_dir, msk_out_dir, msk2_out_dir, label_out_dir = ensure_aug_dirs(
+        group_name=None,
+        use_masks2=use_masks2,
+        use_labels=use_labels,
+    )
 
     count = 0
     for img_file in sorted(imgs):
         base, _ = os.path.splitext(img_file)
-        msk_file = msk_map.get(base)
+        base_norm = normalizar_base(base)
+
+        msk_file = msk_map.get(base_norm)
         if not msk_file:
             print(f"[WARN] (legacy) Máscara não encontrada para {img_file}, pulando.")
             continue
-        msk2_file = msk2_map.get(base) if use_masks2 else None
-        if use_masks2 and not msk2_file:
-            print(f"[WARN] (legacy) mask2 não encontrada para {img_file}, gerando só img+mask.")
+
+        msk2_file = msk2_map.get(base_norm) if use_masks2 else None
+        label_file = label_map.get(base_norm) if use_labels else None
+
+        if use_labels and not label_file and strict_label:
+            print(f"[WARN] (legacy) Label não encontrado para {img_file}, pulando.")
+            continue
+
         try:
             count += augment_pair(
-                os.path.join(ORIG_OLD_IMG, img_file),
-                msk_file,
-                img_out_dir,
-                msk_out_dir,
+                img_path=os.path.join(ORIG_OLD_IMG, img_file),
+                msk_path=msk_file,
+                img_out_dir=img_out_dir,
+                msk_out_dir=msk_out_dir,
                 copies=copies,
                 msk2_path=msk2_file,
-                msk2_out_dir=msk2_out_dir
+                msk2_out_dir=msk2_out_dir,
+                label_path=label_file,
+                label_out_dir=label_out_dir,
+                group_name=None,
             )
         except Exception as e:
             print(f"[ERRO] (legacy) {img_file}: {e}")
+
     print(f"[OK] Legacy → {count} pares gerados.")
     return count
 
-def main(copies=5, groups_csv=None):
+
+def main(copies=5, groups_csv=None, strict_label=False):
     total = 0
+
+    print(f"[INFO] MODELO={MODELO}")
+    print(f"[INFO] dual_head_mask={USE_MASKS2}")
+    print(f"[INFO] dual_head_label={USE_LABELS}")
+
     if os.path.isdir(ORIG_GROUP_ROOT):
         grupos = list_groups(ORIG_GROUP_ROOT)
+
         if groups_csv:
-            # filtra pelos grupos desejados
             want = {g.strip() for g in groups_csv.split(",") if g.strip()}
             grupos = [g for g in grupos if g in want]
             if not grupos:
                 print("[WARN] Nenhum grupo válido encontrado após filtro.")
+
         if not grupos:
             print("[WARN] Nenhum grupo encontrado em original/group. Tentando modo legacy...")
-            total += process_legacy(copies)
+            total += process_legacy(copies, strict_label=strict_label)
         else:
             print(f"Grupos encontrados: {', '.join(grupos)}")
             for g in grupos:
-                total += process_group(g, copies)
+                total += process_group(g, copies, strict_label=strict_label)
     else:
-        # sem estrutura de grupos
-        total += process_legacy(copies)
+        total += process_legacy(copies, strict_label=strict_label)
 
     print(f"\nAugmentation completed! Total: {total} pares gerados.")
 
+
 if __name__ == "__main__":
-    ap = argparse.ArgumentParser(description="Augmentação por grupos (images/masks)")
-    ap.add_argument("--copies", type=int, default=5, help="Número de cópias augmentadas por imagem (default=5).")
-    ap.add_argument("--groups", type=str, default=None, help="Lista de grupos separados por vírgula (ex: chao,erva_cana).")
+    ap = argparse.ArgumentParser(description="Augmentação por grupos com suporte a images/masks/masks2/labels.")
+    ap.add_argument("--copies", type=int, default=5, help="Número de cópias augmentadas por imagem.")
+    ap.add_argument("--groups", type=str, default=None, help="Lista de grupos separados por vírgula.")
+    ap.add_argument("--strict-label", action="store_true", help="Se dual_head_label=true e faltar label, pula o item/grupo.")
     args = ap.parse_args()
-    main(copies=args.copies, groups_csv=args.groups)
+
+    main(copies=args.copies, groups_csv=args.groups, strict_label=args.strict_label)

Python/OAK/datasets/_6_normalize.py

@@ -1,67 +1,86 @@
 #!/usr/bin/env python3
 # -*- coding: utf-8 -*-
 """
-Normaliza/redimensiona imagens e máscaras mantendo a ESTRUTURA POR GRUPO.
+Normaliza/redimensiona imagens, máscaras, masks2 e labels mantendo a ESTRUTURA POR GRUPO.
 
-Entradas (via config.json -> MODELO, RESOLUCAO):
-  - MODELO/dataset/original/group/<grupo>/{images,masks}
-  - MODELO/dataset/augmented/group/<grupo>/{images,masks}
+Entradas via config.json -> camera, resolucao:
+  - MODELO/dataset/original/group/<grupo>/{images,masks,(masks2),(labels)}
+  - MODELO/dataset/augmented/group/<grupo>/{images,masks,(masks2),(labels)}
 
-Saídas (por resolução):
-  - MODELO/dataset/<WxH>/group/<grupo>/{images,masks}
+Saídas por resolução:
+  - MODELO/dataset/<WxH>/group/<grupo>/{images,masks,(masks2),(labels)}
 
-Fallback (modo legado, se não houver "group/"):
-  - original/{images,masks} e augmented/{images,masks} -> <WxH>/{images,masks}
+Fallback legado, se não houver group/:
+  - original/{images,masks,(masks2),(labels)}
+  - augmented/{images,masks,(masks2),(labels)}
+  - saída: <WxH>/{images,masks,(masks2),(labels)}
 
 Conversão de máscara:
-  - Lê máscara RGB e converte para IDs via utils.converter_mask_rgb_para_ids
-  - Ignora classe "ignore" conforme labelmap (usa índice 255 como padrão quando necessário)
+  - Lê máscara RGB e converte para IDs via utils.converter_mask_rgb_para_ids.
+  - Ignore conforme labelmap, usando índice 255 como fallback.
+
+Labels:
+  - Ativado por config['dual_head_label'].
+  - Copia labels .json/.txt para a saída.
+  - Se JSON, atualiza image/mask/mask2/label/base/source/normalized.
+  - Se houver label_id, salva também um .npy com o inteiro para facilitar o Dataset no treino.
 """
+
 import argparse
 import os
 import json
 import cv2
-from typing import Dict, List, Tuple
+from typing import Dict, List, Tuple, Optional
 
 import numpy as np
 from utils import carregar_labelmap_completo, converter_mask_rgb_para_ids
 
-# ⚙️ Configurações
-with open("config_oak.json", "r", encoding="utf-8") as f:
+# ===================== CONFIG =====================
+
+with open("config.json", "r", encoding="utf-8") as f:
     config = json.load(f)
+
 MODELO = config["camera"]
 MODEL_NAME = config["model_name"]
-USE_MASKS2 = config["dual_head"]
-RESOLUCAO = tuple(config["resolucao"])  # [W, H] ou [width, height]
+
+# Compatibilidade:
+# - scripts antigos usam dual_head para masks2
+# - scripts novos podem usar dual_head_mask e dual_head_label separados
+USE_MASKS2 = bool(config.get("dual_head_mask", config.get("dual_head", False)))
+USE_LABELS = bool(config.get("dual_head_label", False))
+
+RESOLUCAO = tuple(config["resolucao"])  # [W, H]
 pasta_base = os.path.join(MODELO, "dataset")
 labelmap_path = os.path.join(pasta_base, "labelmap.txt")
-save_path = os.path.join(MODELO, "backup", config["modelo"], MODEL_NAME)
+mode = "single"
+if USE_MASKS2:
+    mode = "mask2"
+elif USE_LABELS:
+    mode = "label"
+suffix = {
+    "single": "_single",
+    "mask2": "_dual_mask",
+    "label": "_dual_label",
+}[mode]
+save_path = os.path.join(MODELO, "backup", config["modelo"], MODEL_NAME + suffix)
 
-# Dimensões alvo (pode expandir para múltiplas se quiser)
 RESOLUCOES = {
     f"{RESOLUCAO[0]}x{RESOLUCAO[1]}": (RESOLUCAO[0], RESOLUCAO[1]),
 }
 
-# Fontes a processar
-FONTES = ["original", "augmented"]
-
-# Extensões aceitas
 IMG_EXTS = (".jpg", ".jpeg", ".png")
-MSK_EXTS = (".png", ".jpg", ".jpeg")  # preferir .png
-MSK2_EXTS = (".png", ".jpg", ".jpeg")  # idem
+MSK_EXTS = (".png", ".jpg", ".jpeg")
+MSK2_EXTS = (".png", ".jpg", ".jpeg")
+LABEL_EXTS = (".json", ".txt")
 
-# === Acumuladores globais para mean/std dos canais RAW4 ===
-GLOBAL_SUM   = None   # soma por canal
-GLOBAL_SUMSQ = None   # soma dos quadrados por canal
-GLOBAL_PIXELS = 0     # n de pixels por canal (H*W por imagem)
+GLOBAL_SUM = None
+GLOBAL_SUMSQ = None
+GLOBAL_PIXELS = 0
+
+
+# ===================== HELPERS =====================
 
 def infer_ignore_id(ignore_rgb, default_id=255):
-    """
-    Tenta inferir o ID de ignore a partir do valor retornado por carregar_labelmap_completo.
-    - Se for [id] retorna id
-    - Se for (R,G,B) retorna default_id (tipicamente 255)
-    - Se for int, retorna direto
-    """
     if isinstance(ignore_rgb, (list, tuple)):
         if len(ignore_rgb) == 1:
             try:
@@ -74,13 +93,15 @@ def infer_ignore_id(ignore_rgb, default_id=255):
         return ignore_rgb
     return default_id
 
+
 def garantir_dir(p):
     os.makedirs(p, exist_ok=True)
 
+
 def list_groups(root) -> List[str]:
-    """Lista grupos válidos com subpastas images e masks."""
     if not os.path.isdir(root):
         return []
+
     grupos = []
     for name in sorted(os.listdir(root)):
         gdir = os.path.join(root, name)
@@ -90,81 +111,130 @@ def list_groups(root) -> List[str]:
             grupos.append(name)
     return grupos
 
-def map_masks_by_base(msk_dir: str) -> Dict[str, str]:
-    """Retorna {base: caminho_mask}, priorizando .png quando houver múltiplas por base."""
+
+def normalizar_base(stem: str) -> str:
+    sufixos = [
+        "_rgb", "_RGB", "_Rgb",
+        "_image", "_img", "_frame",
+        "_mask", "_masks",
+        "_seg", "_SEG", "_segment", "_segmentacao", "_Segmentacao",
+        "_label", "_labels",
+    ]
+
+    out = stem
+    mudou = True
+    while mudou:
+        mudou = False
+        for sfx in sufixos:
+            if out.endswith(sfx):
+                out = out[: -len(sfx)]
+                mudou = True
+                break
+    return out
+
+
+def map_files_by_base(folder: str, exts: Tuple[str, ...]) -> Dict[str, str]:
     by_base = {}
-    if not os.path.isdir(msk_dir):
+    if not os.path.isdir(folder):
         return by_base
-    for fname in os.listdir(msk_dir):
-        f_lower = fname.lower()
-        if not f_lower.endswith(MSK_EXTS):
+
+    prioridade = {
+        ".json": 0,
+        ".png": 1,
+        ".jpg": 2,
+        ".jpeg": 3,
+        ".txt": 4,
+    }
+
+    for fname in os.listdir(folder):
+        lower = fname.lower()
+        if not lower.endswith(exts):
             continue
-        base, ext = os.path.splitext(fname)
-        cand = os.path.join(msk_dir, fname)
+
+        stem, ext = os.path.splitext(fname)
+        base = normalizar_base(stem)
+        cand = os.path.join(folder, fname)
+
         if base not in by_base:
             by_base[base] = cand
         else:
             cur_ext = os.path.splitext(by_base[base])[1].lower()
-            if cur_ext != ".png" and ext.lower() == ".png":
+            if prioridade.get(ext.lower(), 99) < prioridade.get(cur_ext, 99):
                 by_base[base] = cand
+
     return by_base
 
+
+def map_masks_by_base(msk_dir: str) -> Dict[str, str]:
+    return map_files_by_base(msk_dir, MSK_EXTS)
+
+
 def map_masks2_by_base(msk2_dir: str) -> Dict[str, str]:
-    """Retorna {base: caminho_mask2}, priorizando .png quando houver múltiplas por base."""
-    by_base = {}
-    if not os.path.isdir(msk2_dir):
-        return by_base
-    for fname in os.listdir(msk2_dir):
-        f_lower = fname.lower()
-        if not f_lower.endswith(MSK2_EXTS):
-            continue
-        base, ext = os.path.splitext(fname)
-        cand = os.path.join(msk2_dir, fname)
-        if base not in by_base:
-            by_base[base] = cand
-        else:
-            cur_ext = os.path.splitext(by_base[base])[1].lower()
-            if cur_ext != ".png" and ext.lower() == ".png":
-                by_base[base] = cand
-    return by_base
+    return map_files_by_base(msk2_dir, MSK2_EXTS)
 
-def normalize_pair(caminho_rgb: str, caminho_mask: str, cor_para_id, ignore_id: int, out_img_dir: str, out_msk_dir: str, dim: Tuple[int,int], prefix: str = ""):
-    """Redimensiona e grava a imagem e a máscara (se houver)."""
-    img_rgb = cv2.imread(caminho_rgb)
-    if img_rgb is None:
-        print(f"[!] Erro ao ler imagem: {caminho_rgb}")
-        return False
 
-    # Nome de saída com prefixo para distinguir fonte (ex: original_, augmented_)
-    nome = os.path.basename(caminho_rgb)
-    if prefix:
-        nome_saida_img = f"{prefix}{nome}"
-    else:
-        nome_saida_img = nome
-    nome_saida_msk = nome_saida_img
+def map_labels_by_base(label_dir: str) -> Dict[str, str]:
+    return map_files_by_base(label_dir, LABEL_EXTS)
+
+
+def trocar_ext_para_png(nome: str) -> str:
     for ext in (".jpg", ".jpeg", ".png"):
-        if nome_saida_msk.lower().endswith(ext):
-            nome_saida_msk = nome_saida_msk[: -len(ext)] + ".png"
-            break
+        if nome.lower().endswith(ext):
+            return nome[: -len(ext)] + ".png"
+    return nome + ".png"
 
-    # Redimensiona imagem
-    img_resized = cv2.resize(img_rgb, dim, interpolation=cv2.INTER_AREA)
+
+def safe_rel(path: str, root: str) -> str:
+    try:
+        return os.path.relpath(path, root).replace("\\", "/")
+    except Exception:
+        return str(path).replace("\\", "/")
+
+
+# ===================== NORMALIZAÇÃO =====================
+
+def normalize_pair(
+    caminho_rgb: str,
+    caminho_mask: str,
+    cor_para_id,
+    ignore_id: int,
+    out_img_dir: str,
+    out_msk_dir: str,
+    dim: Tuple[int, int],
+    prefix: str = "",
+):
+    """Redimensiona e grava imagem + máscara ID."""
+    img_bgr = cv2.imread(caminho_rgb)
+    if img_bgr is None:
+        print(f"[!] Erro ao ler imagem: {caminho_rgb}")
+        return False, None, None
+    img_rgb = cv2.cvtColor(img_bgr, cv2.COLOR_BGR2RGB)
+
+    nome = os.path.basename(caminho_rgb)
+    nome_saida_img = f"{prefix}{nome}" if prefix else nome
+    nome_saida_msk = trocar_ext_para_png(nome_saida_img)
+
+    img_resized_rgb = cv2.resize(img_rgb, dim, interpolation=cv2.INTER_AREA)
 
     global GLOBAL_SUM, GLOBAL_SUMSQ, GLOBAL_PIXELS
-    img_float = img_resized.astype(np.float32) / 255.0 # 0-1
+    img_float = img_resized_rgb.astype(np.float32) / 255.0
     h, w, c = img_float.shape
     flat = img_float.reshape(-1, c).astype(np.float64)
-    if (GLOBAL_SUM is None):
+
+    if GLOBAL_SUM is None:
         GLOBAL_SUM = np.zeros(c, dtype=np.float64)
         GLOBAL_SUMSQ = np.zeros(c, dtype=np.float64)
+
     GLOBAL_SUM += flat.sum(axis=0)
     GLOBAL_SUMSQ += (flat ** 2).sum(axis=0)
     GLOBAL_PIXELS += h * w
 
     garantir_dir(out_img_dir)
-    cv2.imwrite(os.path.join(out_img_dir, nome_saida_img), img_resized)
+    out_img_path = os.path.join(out_img_dir, nome_saida_img)
+    img_resized_bgr = cv2.cvtColor(img_resized_rgb, cv2.COLOR_RGB2BGR)
+    cv2.imwrite(out_img_path, img_resized_bgr)
 
-    # Processa e redimensiona máscara (se existir)
+    out_msk_path = None
     if caminho_mask and os.path.isfile(caminho_mask):
         msk_bgr = cv2.imread(caminho_mask, cv2.IMREAD_COLOR)
         if msk_bgr is None:
@@ -173,184 +243,409 @@ def normalize_pair(caminho_rgb: str, caminho_mask: str, cor_para_id, ignore_id:
             msk_rgb = cv2.cvtColor(msk_bgr, cv2.COLOR_BGR2RGB)
             mask_ids = converter_mask_rgb_para_ids(msk_rgb, cor_para_id, ignore_id)
             mask_resized = cv2.resize(mask_ids, dim, interpolation=cv2.INTER_NEAREST)
+
             garantir_dir(out_msk_dir)
-            cv2.imwrite(os.path.join(out_msk_dir, nome_saida_msk), mask_resized)
+            out_msk_path = os.path.join(out_msk_dir, nome_saida_msk)
+            cv2.imwrite(out_msk_path, mask_resized)
+
+    return True, out_img_path, out_msk_path
 
 
-    return True
-
-def normalize_pair_mask2(caminho_rgb: str, caminho_mask2: str,
-                         out_msk2_dir: str, dim: Tuple[int,int], prefix: str = ""):
-    """
-    Redimensiona e grava máscara2 (corredor binário), assumindo que ela já é uma máscara "pronta".
-    - Se for RGB/BGR (3 canais): converte para cinza e faz threshold (0/255) antes de redimensionar.
-    - Se for 1 canal: mantém, faz threshold (0/255).
-    - Redimensiona com INTER_NEAREST.
-    Saída sempre .png com o mesmo nome base/prefixo do arquivo de imagem.
-    """
+def normalize_pair_mask2(
+    caminho_rgb: str,
+    caminho_mask2: str,
+    out_msk2_dir: str,
+    dim: Tuple[int, int],
+    prefix: str = "",
+):
     if not caminho_mask2 or not os.path.isfile(caminho_mask2):
-        return False
+        return False, None
 
     nome = os.path.basename(caminho_rgb)
     nome_saida = f"{prefix}{nome}" if prefix else nome
-    for ext in (".jpg", ".jpeg", ".png"):
-        if nome_saida.lower().endswith(ext):
-            nome_saida = nome_saida[: -len(ext)] + ".png"
-            break
+    nome_saida = trocar_ext_para_png(nome_saida)
 
     m2 = cv2.imread(caminho_mask2, cv2.IMREAD_UNCHANGED)
     if m2 is None:
         print(f"[!] Erro ao ler máscara2: {caminho_mask2}")
-        return False
+        return False, None
 
     if len(m2.shape) == 3:
-        # BGR/RGB -> gray
         m2g = cv2.cvtColor(m2, cv2.COLOR_BGR2GRAY)
     else:
         m2g = m2
 
-    # binariza para 0/255 (evita lixo de compressão)
     _, m2bin = cv2.threshold(m2g, 127, 255, cv2.THRESH_BINARY)
     m2res = cv2.resize(m2bin, dim, interpolation=cv2.INTER_NEAREST)
+
     garantir_dir(out_msk2_dir)
-    cv2.imwrite(os.path.join(out_msk2_dir, nome_saida), m2res)
+    out_msk2_path = os.path.join(out_msk2_dir, nome_saida)
+    cv2.imwrite(out_msk2_path, m2res)
+
+    return True, out_msk2_path
+
+
+def normalize_label(
+    caminho_rgb: str,
+    caminho_label: Optional[str],
+    out_label_dir: Optional[str],
+    out_img_path: Optional[str],
+    out_msk_path: Optional[str],
+    out_msk2_path: Optional[str],
+    source_root: str,
+    output_root: str,
+    grupo: Optional[str],
+    fonte_nome: str,
+    prefix: str = "",
+):
+    """Copia/atualiza label global e salva label_id como .npy quando possível."""
+    if not caminho_label or not out_label_dir or not os.path.isfile(caminho_label):
+        return False
+
+    nome = os.path.basename(caminho_rgb)
+    nome_saida = f"{prefix}{nome}" if prefix else nome
+    base_saida, _ = os.path.splitext(nome_saida)
+
+    label_ext = os.path.splitext(caminho_label)[1].lower()
+    garantir_dir(out_label_dir)
+
+    out_label_path = os.path.join(out_label_dir, base_saida + label_ext)
+    out_label_id_path = os.path.join(out_label_dir, base_saida + ".npy")
+
+    label_id = None
+
+    if label_ext == ".json":
+        try:
+            with open(caminho_label, "r", encoding="utf-8") as f:
+                data = json.load(f)
+        except Exception:
+            data = {}
+
+        label_id = data.get("label_id")
+
+        data["normalized"] = True
+        data["source"] = fonte_nome
+        data["group"] = grupo if grupo is not None else data.get("group")
+        data["base"] = base_saida
+        data["image"] = safe_rel(out_img_path, output_root) if out_img_path else None
+        data["mask"] = safe_rel(out_msk_path, output_root) if out_msk_path else None
+        data["label"] = safe_rel(out_label_path, output_root)
+        data["source_label"] = safe_rel(caminho_label, source_root)
+
+        if out_msk2_path:
+            data["mask2"] = safe_rel(out_msk2_path, output_root)
+
+        with open(out_label_path, "w", encoding="utf-8") as f:
+            json.dump(data, f, indent=2, ensure_ascii=False)
+
+    else:
+        with open(caminho_label, "r", encoding="utf-8") as f:
+            txt = f.read().strip()
+        with open(out_label_path, "w", encoding="utf-8") as f:
+            f.write(txt)
+        label_id = None
+
+    if label_id is not None:
+        try:
+            np.save(out_label_id_path, np.array(int(label_id), dtype=np.int64))
+        except Exception as e:
+            print(f"[WARN] Não consegui salvar label_id npy para {caminho_label}: {e}")
+
     return True
 
-def process_group_root(fonte_root: str, fonte_nome: str, cor_para_id, ignore_id: int, groups_except: str = None):
-    """Processa uma raiz do tipo .../<fonte>/group/ agrupando por cada subpasta de grupo."""
+
+# ===================== PROCESSAMENTO =====================
+
+def process_group_root(
+    fonte_root: str,
+    fonte_nome: str,
+    cor_para_id,
+    ignore_id: int,
+    groups_except: str = "",
+    strict_label: bool = False,
+):
     total = 0
     grupos = list_groups(fonte_root)
     if not grupos:
         return 0
-    
+
     not_want = {g.strip() for g in groups_except.split(",") if g.strip()}
     grupos_desconsiderar = [g for g in grupos if g in not_want]
 
     for nome_res, dim in RESOLUCOES.items():
         out_root = os.path.join(pasta_base, nome_res, "group")
+
         for grupo in grupos:
             if grupo in grupos_desconsiderar:
-                print(f"[WARN] Grupo desconsiderado nao sera processado: {grupo}")
+                print(f"[WARN] Grupo desconsiderado não será processado: {grupo}")
                 continue
+
             in_img_dir = os.path.join(fonte_root, grupo, "images")
             in_msk_dir = os.path.join(fonte_root, grupo, "masks")
             in_msk2_dir = os.path.join(fonte_root, grupo, "masks2")
+            in_label_dir = os.path.join(fonte_root, grupo, "labels")
+
             if not (os.path.isdir(in_img_dir) and os.path.isdir(in_msk_dir)):
-                print(f"[WARN] Grupo inválido (sem images/masks): {grupo}")
+                print(f"[WARN] Grupo inválido sem images/masks: {grupo}")
                 continue
 
             out_img_dir = os.path.join(out_root, grupo, "images")
             out_msk_dir = os.path.join(out_root, grupo, "masks")
+
             usar_masks2 = USE_MASKS2 and os.path.isdir(in_msk2_dir)
+            usar_labels = USE_LABELS and os.path.isdir(in_label_dir)
+
             out_msk2_dir = os.path.join(out_root, grupo, "masks2") if usar_masks2 else None
+            out_label_dir = os.path.join(out_root, grupo, "labels") if usar_labels else None
+
             msk_map = map_masks_by_base(in_msk_dir)
             msk2_map = map_masks2_by_base(in_msk2_dir) if usar_masks2 else {}
+            label_map = map_labels_by_base(in_label_dir) if usar_labels else {}
+
+            if USE_LABELS and not usar_labels:
+                msg = f"[WARN] [{fonte_nome} | {grupo}] dual_head_label=true, mas labels/ não existe."
+                if strict_label:
+                    print(msg + " Pulando grupo.")
+                    continue
+                print(msg + " Seguindo sem labels.")
 
             imgs = [f for f in os.listdir(in_img_dir) if os.path.splitext(f.lower())[1] in IMG_EXTS]
             n = len(imgs)
+
             for i, fname in enumerate(sorted(imgs), 1):
                 base, _ = os.path.splitext(fname)
+                base_norm = normalizar_base(base)
+
                 caminho_rgb = os.path.join(in_img_dir, fname)
-                caminho_mask = msk_map.get(base)
-                caminho_mask2 = msk2_map.get(base) if usar_masks2 else None
-                ok = normalize_pair(
-                    caminho_rgb, caminho_mask, cor_para_id, ignore_id,
-                    out_img_dir, out_msk_dir, dim, prefix=f"{fonte_nome}_"
+                caminho_mask = msk_map.get(base_norm)
+                caminho_mask2 = msk2_map.get(base_norm) if usar_masks2 else None
+                caminho_label = label_map.get(base_norm) if usar_labels else None
+
+                if usar_labels and not caminho_label:
+                    msg = f"[WARN] [{fonte_nome} | {grupo}] label não encontrado para {fname}."
+                    if strict_label:
+                        print(msg + " Pulando item.")
+                        continue
+                    print(msg + " Seguindo sem label.")
+
+                ok, out_img_path, out_msk_path = normalize_pair(
+                    caminho_rgb,
+                    caminho_mask,
+                    cor_para_id,
+                    ignore_id,
+                    out_img_dir,
+                    out_msk_dir,
+                    dim,
+                    prefix=f"{fonte_nome}_",
                 )
+
+                out_msk2_path = None
                 if usar_masks2 and out_msk2_dir:
                     if not caminho_mask2:
-                        print(f"[WARN] [{fonte_nome} | {grupo}] masks2 existe, mas não achei mask2 p/ {fname} (vou seguir).")
+                        print(f"[WARN] [{fonte_nome} | {grupo}] masks2 existe, mas não achei mask2 p/ {fname}.")
                     else:
-                        normalize_pair_mask2(
-                            caminho_rgb, caminho_mask2,
-                            out_msk2_dir, dim, prefix=f"{fonte_nome}_"
+                        _, out_msk2_path = normalize_pair_mask2(
+                            caminho_rgb,
+                            caminho_mask2,
+                            out_msk2_dir,
+                            dim,
+                            prefix=f"{fonte_nome}_",
                         )
+
+                if usar_labels and caminho_label and out_label_dir:
+                    normalize_label(
+                        caminho_rgb=caminho_rgb,
+                        caminho_label=caminho_label,
+                        out_label_dir=out_label_dir,
+                        out_img_path=out_img_path,
+                        out_msk_path=out_msk_path,
+                        out_msk2_path=out_msk2_path,
+                        source_root=fonte_root,
+                        output_root=os.path.join(pasta_base, nome_res),
+                        grupo=grupo,
+                        fonte_nome=fonte_nome,
+                        prefix=f"{fonte_nome}_",
+                    )
+
                 if ok:
                     total += 1
+
                 print(f"[{fonte_nome} | {grupo} | {nome_res}] {i}/{n} → {fname}")
+
     return total
 
-def process_legacy_root(legacy_img: str, legacy_msk: str, fonte_nome: str, cor_para_id, ignore_id: int):
-    """Processa estrutura legado (sem grupos)."""
+
+def process_legacy_root(
+    legacy_img: str,
+    legacy_msk: str,
+    fonte_nome: str,
+    cor_para_id,
+    ignore_id: int,
+    strict_label: bool = False,
+):
     if not (os.path.isdir(legacy_img) and os.path.isdir(legacy_msk)):
         return 0
 
     total = 0
+
     for nome_res, dim in RESOLUCOES.items():
         out_img_dir = os.path.join(pasta_base, nome_res, "images")
         out_msk_dir = os.path.join(pasta_base, nome_res, "masks")
-        legacy_msk2 = os.path.join(os.path.dirname(legacy_msk), "masks2")
+
+        legacy_root = os.path.dirname(legacy_msk)
+        legacy_msk2 = os.path.join(legacy_root, "masks2")
+        legacy_labels = os.path.join(legacy_root, "labels")
+
         usar_masks2 = USE_MASKS2 and os.path.isdir(legacy_msk2)
+        usar_labels = USE_LABELS and os.path.isdir(legacy_labels)
+
         out_msk2_dir = os.path.join(pasta_base, nome_res, "masks2") if usar_masks2 else None
+        out_label_dir = os.path.join(pasta_base, nome_res, "labels") if usar_labels else None
 
         msk_map = map_masks_by_base(legacy_msk)
         msk2_map = map_masks2_by_base(legacy_msk2) if usar_masks2 else {}
+        label_map = map_labels_by_base(legacy_labels) if usar_labels else {}
+
+        if USE_LABELS and not usar_labels:
+            msg = f"[WARN] [{fonte_nome} | legacy] dual_head_label=true, mas labels/ não existe."
+            if strict_label:
+                print(msg + " Pulando legacy.")
+                return total
+            print(msg + " Seguindo sem labels.")
+
         imgs = [f for f in os.listdir(legacy_img) if os.path.splitext(f.lower())[1] in IMG_EXTS]
         n = len(imgs)
+
         for i, fname in enumerate(sorted(imgs), 1):
             base, _ = os.path.splitext(fname)
+            base_norm = normalizar_base(base)
+
             caminho_rgb = os.path.join(legacy_img, fname)
-            caminho_mask = msk_map.get(base)
-            caminho_mask2 = msk2_map.get(base) if usar_masks2 else None
-            ok = normalize_pair(
-                caminho_rgb, caminho_mask, cor_para_id, ignore_id,
-                out_img_dir, out_msk_dir, dim, prefix=f"{fonte_nome}_"
+            caminho_mask = msk_map.get(base_norm)
+            caminho_mask2 = msk2_map.get(base_norm) if usar_masks2 else None
+            caminho_label = label_map.get(base_norm) if usar_labels else None
+
+            if usar_labels and not caminho_label:
+                msg = f"[WARN] [{fonte_nome} | legacy] label não encontrado para {fname}."
+                if strict_label:
+                    print(msg + " Pulando item.")
+                    continue
+                print(msg + " Seguindo sem label.")
+
+            ok, out_img_path, out_msk_path = normalize_pair(
+                caminho_rgb,
+                caminho_mask,
+                cor_para_id,
+                ignore_id,
+                out_img_dir,
+                out_msk_dir,
+                dim,
+                prefix=f"{fonte_nome}_",
             )
+
+            out_msk2_path = None
             if usar_masks2 and out_msk2_dir:
                 if not caminho_mask2:
-                    print(f"[WARN] [{fonte_nome} | legacy] masks2 existe, mas não achei mask2 p/ {fname} (vou seguir).")
+                    print(f"[WARN] [{fonte_nome} | legacy] masks2 existe, mas não achei mask2 p/ {fname}.")
                 else:
-                    normalize_pair_mask2(
-                        caminho_rgb, caminho_mask2,
-                        out_msk2_dir, dim, prefix=f"{fonte_nome}_"
+                    _, out_msk2_path = normalize_pair_mask2(
+                        caminho_rgb,
+                        caminho_mask2,
+                        out_msk2_dir,
+                        dim,
+                        prefix=f"{fonte_nome}_",
                     )
+
+            if usar_labels and caminho_label and out_label_dir:
+                normalize_label(
+                    caminho_rgb=caminho_rgb,
+                    caminho_label=caminho_label,
+                    out_label_dir=out_label_dir,
+                    out_img_path=out_img_path,
+                    out_msk_path=out_msk_path,
+                    out_msk2_path=out_msk2_path,
+                    source_root=legacy_root,
+                    output_root=os.path.join(pasta_base, nome_res),
+                    grupo=None,
+                    fonte_nome=fonte_nome,
+                    prefix=f"{fonte_nome}_",
+                )
+
             if ok:
                 total += 1
+
             print(f"[{fonte_nome} | legacy | {nome_res}] {i}/{n} → {fname}")
+
     return total
 
+
+# ===================== MAIN =====================
+
 def main(args):
-    # === Labelmap ===
-    # Espera tupla na ordem: (cor_para_id, colormap_rgb, id_para_nome, ignore_rgb)
     cor_para_id, _colormap_rgb, _id_para_nome, ignore_rgb = carregar_labelmap_completo(labelmap_path)
     ignore_id = infer_ignore_id(ignore_rgb, default_id=255)
-    print(cor_para_id, _colormap_rgb, _id_para_nome)
+
+    print(f"[INFO] cor_para_id: {cor_para_id}")
+    print(f"[INFO] classes: {_id_para_nome}")
+    print(f"[INFO] dual_head_mask/masks2: {USE_MASKS2}")
+    print(f"[INFO] dual_head_label/labels: {USE_LABELS}")
 
     total_geral = 0
-    # === ORIGINAL ===
+
     orig_group_root = os.path.join(pasta_base, "original", "group")
     if os.path.isdir(orig_group_root):
-        total_geral += process_group_root(orig_group_root, "original", cor_para_id, ignore_id, groups_except=args.groups_except)
+        total_geral += process_group_root(
+            orig_group_root,
+            "original",
+            cor_para_id,
+            ignore_id,
+            groups_except=args.groups_except,
+            strict_label=args.strict_label,
+        )
     else:
         legacy_img = os.path.join(pasta_base, "original", "images")
         legacy_msk = os.path.join(pasta_base, "original", "masks")
-        total_geral += process_legacy_root(legacy_img, legacy_msk, "original", cor_para_id, ignore_id)
+        total_geral += process_legacy_root(
+            legacy_img,
+            legacy_msk,
+            "original",
+            cor_para_id,
+            ignore_id,
+            strict_label=args.strict_label,
+        )
 
-    # === AUGMENTED ===
     aug_group_root = os.path.join(pasta_base, "augmented", "group")
     if os.path.isdir(aug_group_root):
-        total_geral += process_group_root(aug_group_root, "augmented", cor_para_id, ignore_id, groups_except=args.groups_except)
+        total_geral += process_group_root(
+            aug_group_root,
+            "augmented",
+            cor_para_id,
+            ignore_id,
+            groups_except=args.groups_except,
+            strict_label=args.strict_label,
+        )
     else:
         legacy_img = os.path.join(pasta_base, "augmented", "images")
         legacy_msk = os.path.join(pasta_base, "augmented", "masks")
-        total_geral += process_legacy_root(legacy_img, legacy_msk, "augmented", cor_para_id, ignore_id)
+        total_geral += process_legacy_root(
+            legacy_img,
+            legacy_msk,
+            "augmented",
+            cor_para_id,
+            ignore_id,
+            strict_label=args.strict_label,
+        )
 
     print(f"\n✅ Concluído! Total normalizados: {total_geral}")
 
-    # === calcula mean/std globais e salva em JSON ===
     global GLOBAL_SUM, GLOBAL_SUMSQ, GLOBAL_PIXELS
     if GLOBAL_SUM is not None and GLOBAL_PIXELS > 0:
-        # média e variância por canal
-        mean = (GLOBAL_SUM / GLOBAL_PIXELS)
-        var  = (GLOBAL_SUMSQ / GLOBAL_PIXELS) - mean**2
-        std  = np.sqrt(np.maximum(var, 1e-6))
+        mean = GLOBAL_SUM / GLOBAL_PIXELS
+        var = (GLOBAL_SUMSQ / GLOBAL_PIXELS) - mean ** 2
+        std = np.sqrt(np.maximum(var, 1e-6))
 
-        # Converte para list pra salvar em JSON
         mean_list = mean.tolist()
-        std_list  = std.tolist()
-
-        # Se quiser, você pode nomear os canais explicitamente
-        # dependendo da convenção do raw4:
+        std_list = std.tolist()
         channel_names = ["R", "G", "B"]
 
         stats = {
@@ -369,10 +664,21 @@ def main(args):
         print(f"  mean: {mean_list}")
         print(f"  std : {std_list}")
     else:
-        print("⚠️ Nenhum RAW processado, não há stats para salvar.")
+        print("⚠️ Nenhuma imagem processada, não há stats para salvar.")
+
 
 if __name__ == "__main__":
-    ap = argparse.ArgumentParser(description="Augmentação por grupos (images/masks)")
-    ap.add_argument("--groups-except", type=str, default="", help="Lista de grupos para nao usar, separados por vírgula (ex: chao,erva_cana).")
+    ap = argparse.ArgumentParser(description="Normalização por grupos images/masks/masks2/labels.")
+    ap.add_argument(
+        "--groups-except",
+        type=str,
+        default="",
+        help="Lista de grupos para não usar, separados por vírgula. Ex: chao,erva_cana",
+    )
+    ap.add_argument(
+        "--strict-label",
+        action="store_true",
+        help="Se dual_head_label=true e faltar label, pula item/grupo.",
+    )
     args = ap.parse_args()
     main(args)

Python/OAK/datasets/_7_split.py

@@ -1,72 +1,88 @@
 #!/usr/bin/env python3
 # -*- coding: utf-8 -*-
 """
-Split estratificado por GRUPO com **val/test só do ORIGINAL** e
-garantia de NÃO VAZAMENTO entre splits (mesma família não cruza splits).
+Split estratificado por GRUPO com val/test só do ORIGINAL e garantia de NÃO VAZAMENTO.
 
 Lê de:
-  MODELO/dataset/<WxH>/group/<grupo>/{images,masks}
+  MODELO/dataset/<WxH>/group/<grupo>/{images,masks,(masks2),(labels)}
 
 Escreve em:
-  MODELO/dataset/split/<split>/group/<grupo>/{images,masks}
+  MODELO/dataset/split/<split>/group/<grupo>/{images,masks,(masks2),(labels)}
 
 Definições:
-- "Família" = todas as variações da MESMA base original:
-    original_<base>.* e augmented_<base>_aug_XX.*
-- Val/Test: só **original_<base>** (sem augmented)
-- Train: original_<base> **e** todos augmented_<base>_aug_XX
+- Família = todas as variações da mesma base original:
+    original_<base>.*
+    augmented_<base>_aug_XX.*
+- Val/Test: somente original_<base>.
+- Train: original_<base> + todos augmented_<base>_aug_XX.
 
-Se não houver prefixos (legado), cai para o comportamento antigo (sem família),
-mas ainda evita colocar augmented em val/test se detectar sufixo "_aug_XX".
+Labels:
+- Ativados por config['dual_head_label'].
+- Copia labels .json/.txt e .npy quando existirem.
+- O pareamento é feito pelo mesmo base name da imagem.
 
 Uso:
-  python _7_split_grouped_noleak.py
-  python _7_split_grouped_noleak.py --train 0.7 --val 0.29 --test 0.01 --seed 42
-  python _7_split_grouped_noleak.py --min-train 1 --min-val 1 --min-test 0
-  python _7_split_grouped_noleak.py --modelo OAK-1-Lite-W --resolucao 640x384
+  python _7_split_grouped_noleak_with_labels.py
+  python _7_split_grouped_noleak_with_labels.py --train 0.7 --val 0.29 --test 0.01 --seed 42
+  python _7_split_grouped_noleak_with_labels.py --strict-label
+  python _7_split_grouped_noleak_with_labels.py --cap-train-families "navegavel:300,naonavegavel_navegavel:800"
 """
+
 import os
 import re
 import json
 import shutil
 import random
 import argparse
+from typing import Dict, List, Optional
 
-# ⚙️ Configurações
-with open("config_oak.json", "r", encoding="utf-8") as f:
+
+# ===================== CONFIG =====================
+
+with open("config.json", "r", encoding="utf-8") as f:
     config = json.load(f)
+
 MODELO = config.get("camera")
-USE_MASKS2 = config.get("dual_head", False)
+USE_MASKS2 = bool(config.get("dual_head_mask", config.get("dual_head", False)))
+USE_LABELS = bool(config.get("dual_head_label", False))
 RESOLUCAO = tuple(config.get("resolucao"))
 
-# Pastas
 pasta_origem = os.path.join(MODELO, "dataset", f"{RESOLUCAO[0]}x{RESOLUCAO[1]}", "group")
 pasta_destino = os.path.join(MODELO, "dataset", "split")
 
 IMG_EXTS = (".jpg", ".jpeg", ".png")
-MSK_EXT  = ".png"  # máscaras normalizadas em PNG (recomendado)
+MSK_EXT = ".png"
+LABEL_EXTS = (".json", ".txt", ".npy")
 
-# Regex para identificar famílias
-RE_ORIGINAL_PREFIX  = re.compile(r'^original_(.+)$', re.IGNORECASE)
-RE_AUGMENTED_FAMILY = re.compile(r'^augmented_(.+?)(?:_aug_\d+)?$', re.IGNORECASE)
-RE_AUG_SUFFIX       = re.compile(r'_aug_\d+$', re.IGNORECASE)
+RE_ORIGINAL_PREFIX = re.compile(r"^original_(.+)$", re.IGNORECASE)
+RE_AUGMENTED_FAMILY = re.compile(r"^augmented_(.+?)(?:_aug_\d+)?$", re.IGNORECASE)
+RE_AUG_SUFFIX = re.compile(r"_aug_\d+$", re.IGNORECASE)
+
+
+# ===================== HELPERS =====================
 
 def garantir(p):
     os.makedirs(p, exist_ok=True)
 
+
 def lista_grupos(root):
     if not os.path.isdir(root):
         return []
+
     out = []
     for g in sorted(os.listdir(root)):
         gdir = os.path.join(root, g)
-        if not os.path.isdir(gdir): continue
+        if not os.path.isdir(gdir):
+            continue
         if os.path.isdir(os.path.join(gdir, "images")) and os.path.isdir(os.path.join(gdir, "masks")):
             out.append(g)
     return out
 
+
 def listar_imagens(img_dir):
-    if not os.path.isdir(img_dir): return []
+    if not os.path.isdir(img_dir):
+        return []
+
     fs = []
     for f in os.listdir(img_dir):
         ext = os.path.splitext(f.lower())[1]
@@ -74,19 +90,27 @@ def listar_imagens(img_dir):
             fs.append(f)
     return sorted(fs)
 
+
 def mask_from_image_name(img_name):
     base, _ = os.path.splitext(img_name)
     return base + MSK_EXT
 
+
 def mask2_from_image_name(img_name):
     base, _ = os.path.splitext(img_name)
-    return base + MSK_EXT  # masks2 também normalizadas em PNG no normalize
+    return base + MSK_EXT
+
+
+def label_candidates_from_image_name(img_name):
+    base, _ = os.path.splitext(img_name)
+    return [base + ext for ext in LABEL_EXTS]
+
 
 def classify_source_and_family(filename_no_ext):
     """
     Retorna (source, family_key)
-      source ∈ {"original", "augmented", "unknown"}
-      family_key = base associada ao original (sem prefixo/sufixos), ex: "foo_001"
+      source ∈ {original, augmented, unknown}
+      family_key = base original sem prefixo/sufixo.
     """
     m = RE_ORIGINAL_PREFIX.match(filename_no_ext)
     if m:
@@ -96,46 +120,66 @@ def classify_source_and_family(filename_no_ext):
     if m:
         return "augmented", m.group(1)
 
-    # legado: tenta deduzir se é augmented por sufixo, e família é o próprio nome sem sufixo
     if RE_AUG_SUFFIX.search(filename_no_ext):
         fam = RE_AUG_SUFFIX.sub("", filename_no_ext)
         return "augmented", fam
 
     return "unknown", filename_no_ext
 
-def build_family_index(img_dir, msk_dir):
+
+def has_any_label(label_dir: str, img_name: str) -> bool:
+    if not label_dir or not os.path.isdir(label_dir):
+        return False
+    return any(os.path.exists(os.path.join(label_dir, cand)) for cand in label_candidates_from_image_name(img_name))
+
+
+def build_family_index(img_dir, msk_dir, label_dir=None, require_label=False):
     """
-    Constroi índice de famílias a partir de img_dir/msk_dir.
-    Retorna: dict family -> {"original": str|None, "augmented": [str], "all": [str]}
-    (strings são NOMES DE ARQUIVO, não paths completos; assumem que a máscara existe)
+    Constrói índice de famílias a partir de img_dir/msk_dir/labels.
+
+    Retorna:
+      dict family -> {
+          original: str|None,
+          augmented: [str],
+          all: [str]
+      }
+
+    Os nomes são arquivos de imagem.
     """
     familias = {}
     imgs = listar_imagens(img_dir)
+
     for img_name in imgs:
-        base_no_ext, ext = os.path.splitext(img_name)
+        base_no_ext, _ = os.path.splitext(img_name)
         mask_name = mask_from_image_name(img_name)
+
         if not os.path.exists(os.path.join(msk_dir, mask_name)):
-            continue  # garante pareamento
+            continue
+
+        if require_label and not has_any_label(label_dir, img_name):
+            continue
 
         source, fam = classify_source_and_family(base_no_ext)
         d = familias.setdefault(fam, {"original": None, "augmented": [], "all": []})
         d["all"].append(img_name)
+
         if source == "original":
             d["original"] = img_name
         elif source == "augmented":
             d["augmented"].append(img_name)
         else:
-            # trata como original desconhecido para não perder dado
             if d["original"] is None:
                 d["original"] = img_name
             else:
                 d["augmented"].append(img_name)
+
     return familias
 
+
 def allocate_counts(n, p_train, p_val, p_test, min_train, min_val, min_test):
     n_train = int(round(n * p_train))
-    n_val   = int(round(n * p_val))
-    n_test  = n - n_train - n_val
+    n_val = int(round(n * p_val))
+    n_test = n - n_train - n_val
 
     if n_test < 0:
         excesso = -n_test
@@ -151,8 +195,8 @@ def allocate_counts(n, p_train, p_val, p_test, min_train, min_val, min_test):
     min_sum = min_train + min_val + min_test
     if n >= min_sum:
         n_train = max(n_train, min_train)
-        n_val   = max(n_val,   min_val)
-        n_test  = max(n_test,  min_test)
+        n_val = max(n_val, min_val)
+        n_test = max(n_test, min_test)
 
         total = n_train + n_val + n_test
         while total > n:
@@ -165,6 +209,7 @@ def allocate_counts(n, p_train, p_val, p_test, min_train, min_val, min_test):
             else:
                 break
             total = n_train + n_val + n_test
+
         while total < n:
             if n_train - min_train <= n_val - min_val:
                 n_train += 1
@@ -177,11 +222,9 @@ def allocate_counts(n, p_train, p_val, p_test, min_train, min_val, min_test):
         n_val = max(0, min(resto, min_val))
         n_test = max(0, resto - n_val)
 
-    # ajuste final
     diff = n - (n_train + n_val + n_test)
     if diff != 0:
         if diff > 0:
-            # adiciona em train, depois val
             take = min(diff, n - n_train)
             n_train += take
             diff -= take
@@ -189,7 +232,6 @@ def allocate_counts(n, p_train, p_val, p_test, min_train, min_val, min_test):
                 n_val += diff
         else:
             diff = -diff
-            # tira de test, depois val
             take = min(diff, n_test)
             n_test -= take
             diff -= take
@@ -198,38 +240,113 @@ def allocate_counts(n, p_train, p_val, p_test, min_train, min_val, min_test):
 
     return n_train, n_val, n_test
 
-def copiar(nomes, src_img_dir, src_msk_dir, dst_img_dir, dst_msk_dir, src_msk2_dir=None, dst_msk2_dir=None):
-    garantir(dst_img_dir); garantir(dst_msk_dir)
+
+def copiar_labels_para_item(nome, src_label_dir, dst_label_dir):
+    if not src_label_dir or not dst_label_dir or not os.path.isdir(src_label_dir):
+        return 0
+
+    garantir(dst_label_dir)
+    copied = 0
+
+    for cand in label_candidates_from_image_name(nome):
+        src = os.path.join(src_label_dir, cand)
+        if not os.path.exists(src):
+            continue
+        shutil.copy2(src, os.path.join(dst_label_dir, cand))
+        copied += 1
+
+    return copied
+
+
+def copiar(
+    nomes,
+    src_img_dir,
+    src_msk_dir,
+    dst_img_dir,
+    dst_msk_dir,
+    src_msk2_dir=None,
+    dst_msk2_dir=None,
+    src_label_dir=None,
+    dst_label_dir=None,
+    strict_label=False,
+):
+    garantir(dst_img_dir)
+    garantir(dst_msk_dir)
+
     use_msk2 = bool(src_msk2_dir and dst_msk2_dir and os.path.isdir(src_msk2_dir))
+    use_labels = bool(src_label_dir and dst_label_dir and os.path.isdir(src_label_dir))
+
     if use_msk2:
         garantir(dst_msk2_dir)
+    if use_labels:
+        garantir(dst_label_dir)
+
     moved = 0
+    skipped_no_label = 0
+
     for nome in nomes:
         mask_name = mask_from_image_name(nome)
         src_img = os.path.join(src_img_dir, nome)
         src_msk = os.path.join(src_msk_dir, mask_name)
+
         if not (os.path.exists(src_img) and os.path.exists(src_msk)):
             continue
+
+        if strict_label and use_labels and not has_any_label(src_label_dir, nome):
+            skipped_no_label += 1
+            continue
+
         shutil.copy2(src_img, os.path.join(dst_img_dir, nome))
         shutil.copy2(src_msk, os.path.join(dst_msk_dir, mask_name))
+
         if use_msk2:
             m2_name = mask2_from_image_name(nome)
             src_m2 = os.path.join(src_msk2_dir, m2_name)
             if os.path.exists(src_m2):
                 shutil.copy2(src_m2, os.path.join(dst_msk2_dir, m2_name))
+
+        if use_labels:
+            copied = copiar_labels_para_item(nome, src_label_dir, dst_label_dir)
+            if strict_label and copied == 0:
+                skipped_no_label += 1
+                continue
+
         moved += 1
+
+    if skipped_no_label > 0:
+        print(f"[WARN] {skipped_no_label} itens pulados por falta de label.")
+
     return moved
 
-def split_group(group_name, p_train, p_val, p_test, seed, mins, caps_map=None):
+
+# ===================== SPLIT =====================
+
+def split_group(group_name, p_train, p_val, p_test, seed, mins, caps_map=None, strict_label=False):
     src_img_dir = os.path.join(pasta_origem, group_name, "images")
     src_msk_dir = os.path.join(pasta_origem, group_name, "masks")
     src_msk2_dir = os.path.join(pasta_origem, group_name, "masks2")
-    use_msk2 = USE_MASKS2 and os.path.isdir(src_msk2_dir)
+    src_label_dir = os.path.join(pasta_origem, group_name, "labels")
+
+    use_msk2 = USE_MASKS2 and os.path.isdir(src_msk2_dir)
+    use_labels = USE_LABELS and os.path.isdir(src_label_dir)
+
+    if USE_LABELS and not use_labels:
+        msg = f"[{group_name}] dual_head_label=true, mas labels/ não existe."
+        if strict_label:
+            print(f"[WARN] {msg} Pulando grupo.")
+            return {"train": 0, "val": 0, "test": 0, "familias": 0}
+        print(f"[WARN] {msg} Split seguirá sem copiar labels.")
+
+    familias = build_family_index(
+        src_img_dir,
+        src_msk_dir,
+        label_dir=src_label_dir if use_labels else None,
+        require_label=bool(strict_label and use_labels),
+    )
 
-    familias = build_family_index(src_img_dir, src_msk_dir)
-    # apenas famílias que têm ORIGINAL para participar de val/test
     familias_originais = [fam for fam, d in familias.items() if d["original"] is not None]
     total_familias = len(familias_originais)
+
     if total_familias == 0:
         print(f"[{group_name}] 0 famílias com original, pulando.")
         return {"train": 0, "val": 0, "test": 0, "familias": 0}
@@ -238,81 +355,145 @@ def split_group(group_name, p_train, p_val, p_test, seed, mins, caps_map=None):
     rng.shuffle(familias_originais)
 
     n_tr, n_va, n_te = allocate_counts(
-        total_familias, p_train, p_val, p_test,
-        mins["train"], mins["val"], mins["test"]
+        total_familias,
+        p_train,
+        p_val,
+        p_test,
+        mins["train"],
+        mins["val"],
+        mins["test"],
     )
 
     fam_train = set(familias_originais[:n_tr])
-    fam_val   = set(familias_originais[n_tr:n_tr+n_va])
-    fam_test  = set(familias_originais[n_tr+n_va: n_tr+n_va+n_te])
+    fam_val = set(familias_originais[n_tr : n_tr + n_va])
+    fam_test = set(familias_originais[n_tr + n_va : n_tr + n_va + n_te])
 
-    # --- CAP por grupo (apenas no TRAIN) ---
     if caps_map and group_name in caps_map:
         cap = caps_map[group_name]
         if len(fam_train) > cap:
             fam_list = list(fam_train)
-            rng.shuffle(fam_list)  # usa o rng já criado com seed
-            kept     = set(fam_list[:cap])
-            dropped  = set(fam_list[cap:])
+            rng.shuffle(fam_list)
+            kept = set(fam_list[:cap])
+            dropped = set(fam_list[cap:])
             fam_train = kept
-            print(f"[{group_name}] cap-train-families={cap} → mantidas {len(kept)} famílias, descartadas {len(dropped)} do TRAIN")
+            print(
+                f"[{group_name}] cap-train-families={cap} → "
+                f"mantidas {len(kept)} famílias, descartadas {len(dropped)} do TRAIN"
+            )
 
-    # listas de nomes por split (imagens)
     nomes_train, nomes_val, nomes_test = [], [], []
 
     for fam, d in familias.items():
         if fam in fam_train:
-            # train recebe original + todos augmented
             if d["original"]:
                 nomes_train.append(d["original"])
             if d["augmented"]:
                 nomes_train.extend(d["augmented"])
         elif fam in fam_val:
-            # val recebe somente original
             if d["original"]:
                 nomes_val.append(d["original"])
         elif fam in fam_test:
-            # test recebe somente original
             if d["original"]:
                 nomes_test.append(d["original"])
-        else:
-            # famílias sem original (não devem cair aqui) ficam fora
-            pass
 
-    # dest dirs
     dest_train_img = os.path.join(pasta_destino, "train", "group", group_name, "images")
     dest_train_msk = os.path.join(pasta_destino, "train", "group", group_name, "masks")
-    dest_val_img   = os.path.join(pasta_destino, "val",   "group", group_name, "images")
-    dest_val_msk   = os.path.join(pasta_destino, "val",   "group", group_name, "masks")
-    dest_test_img  = os.path.join(pasta_destino, "test",  "group", group_name, "images")
-    dest_test_msk  = os.path.join(pasta_destino, "test",  "group", group_name, "masks")
+    dest_val_img = os.path.join(pasta_destino, "val", "group", group_name, "images")
+    dest_val_msk = os.path.join(pasta_destino, "val", "group", group_name, "masks")
+    dest_test_img = os.path.join(pasta_destino, "test", "group", group_name, "images")
+    dest_test_msk = os.path.join(pasta_destino, "test", "group", group_name, "masks")
 
     dest_train_msk2 = os.path.join(pasta_destino, "train", "group", group_name, "masks2") if use_msk2 else None
-    dest_val_msk2   = os.path.join(pasta_destino, "val",   "group", group_name, "masks2") if use_msk2 else None
-    dest_test_msk2  = os.path.join(pasta_destino, "test",  "group", group_name, "masks2") if use_msk2 else None
+    dest_val_msk2 = os.path.join(pasta_destino, "val", "group", group_name, "masks2") if use_msk2 else None
+    dest_test_msk2 = os.path.join(pasta_destino, "test", "group", group_name, "masks2") if use_msk2 else None
 
-    m_train = copiar(nomes_train, src_img_dir, src_msk_dir, dest_train_img, dest_train_msk, src_msk2_dir, dest_train_msk2)
-    m_val   = copiar(nomes_val,   src_img_dir, src_msk_dir, dest_val_img,   dest_val_msk,   src_msk2_dir, dest_val_msk2)
-    m_test  = copiar(nomes_test,  src_img_dir, src_msk_dir, dest_test_img,  dest_test_msk,  src_msk2_dir, dest_test_msk2)
+    dest_train_label = os.path.join(pasta_destino, "train", "group", group_name, "labels") if use_labels else None
+    dest_val_label = os.path.join(pasta_destino, "val", "group", group_name, "labels") if use_labels else None
+    dest_test_label = os.path.join(pasta_destino, "test", "group", group_name, "labels") if use_labels else None
 
-    print(f"[{group_name}] famílias={total_familias}  →  train(imgs)={m_train}, val(imgs)={m_val}, test(imgs)={m_test}")
+    m_train = copiar(
+        nomes_train,
+        src_img_dir,
+        src_msk_dir,
+        dest_train_img,
+        dest_train_msk,
+        src_msk2_dir,
+        dest_train_msk2,
+        src_label_dir,
+        dest_train_label,
+        strict_label=strict_label,
+    )
+    m_val = copiar(
+        nomes_val,
+        src_img_dir,
+        src_msk_dir,
+        dest_val_img,
+        dest_val_msk,
+        src_msk2_dir,
+        dest_val_msk2,
+        src_label_dir,
+        dest_val_label,
+        strict_label=strict_label,
+    )
+    m_test = copiar(
+        nomes_test,
+        src_img_dir,
+        src_msk_dir,
+        dest_test_img,
+        dest_test_msk,
+        src_msk2_dir,
+        dest_test_msk2,
+        src_label_dir,
+        dest_test_label,
+        strict_label=strict_label,
+    )
+
+    print(
+        f"[{group_name}] famílias={total_familias} → "
+        f"train(imgs)={m_train}, val(imgs)={m_val}, test(imgs)={m_test}"
+    )
     return {"train": m_train, "val": m_val, "test": m_test, "familias": total_familias}
 
-def main():
-    ap = argparse.ArgumentParser(description="Split estratificado por grupo SEM vazamento (val/test só original).")
-    ap.add_argument("--train", type=float, default=0.70, help="Proporção de treino (default=0.70).")
-    ap.add_argument("--val",   type=float, default=0.29, help="Proporção de validação (default=0.29).")
-    ap.add_argument("--test",  type=float, default=0.01, help="Proporção de teste (default=0.01).")
-    ap.add_argument("--seed",  type=int,   default=42,   help="Seed do embaralhamento (default=42).")
 
-    ap.add_argument("--min-train", type=int, default=1, help="Mínimo de FAMÍLIAS por grupo em train (default=1).")
-    ap.add_argument("--min-val",   type=int, default=1, help="Mínimo de FAMÍLIAS por grupo em val (default=1).")
-    ap.add_argument("--min-test",  type=int, default=0, help="Mínimo de FAMÍLIAS por grupo em test (default=0).")
+# ===================== MAIN =====================
+
+def parse_cap_map(s):
+    caps = {}
+    if not s:
+        return caps
+    for item in s.split(","):
+        if not item.strip():
+            continue
+        k, v = item.strip().split(":")
+        caps[k.strip()] = int(v)
+    return caps
+
+
+def main():
+    ap = argparse.ArgumentParser(description="Split estratificado por grupo sem vazamento, com suporte a labels.")
+    ap.add_argument("--train", type=float, default=0.70, help="Proporção de treino.")
+    ap.add_argument("--val", type=float, default=0.29, help="Proporção de validação.")
+    ap.add_argument("--test", type=float, default=0.01, help="Proporção de teste.")
+    ap.add_argument("--seed", type=int, default=42, help="Seed do embaralhamento.")
+
+    ap.add_argument("--min-train", type=int, default=1, help="Mínimo de famílias por grupo em train.")
+    ap.add_argument("--min-val", type=int, default=1, help="Mínimo de famílias por grupo em val.")
+    ap.add_argument("--min-test", type=int, default=0, help="Mínimo de famílias por grupo em test.")
 
     ap.add_argument("--modelo", type=str, default=None, help="Sobrescreve MODELO do config.json.")
-    ap.add_argument("--resolucao", type=str, default=None, help="Sobrescreve resolução no formato WxH (ex: 640x480).")
+    ap.add_argument("--resolucao", type=str, default=None, help="Sobrescreve resolução no formato WxH.")
 
-    ap.add_argument("--cap-train-families", type=str, default="", help="Mapa 'grupo:cap,...' p/ limitar número de FAMÍLIAS no TRAIN. Ex.: 'chao:350'")
+    ap.add_argument(
+        "--cap-train-families",
+        type=str,
+        default="",
+        help="Mapa grupo:cap para limitar famílias no TRAIN. Ex: 'navegavel:350'",
+    )
+    ap.add_argument(
+        "--strict-label",
+        action="store_true",
+        help="Se dual_head_label=true e faltar label, pula item/grupo.",
+    )
 
     args = ap.parse_args()
 
@@ -326,14 +507,6 @@ def main():
     else:
         resolucao = RESOLUCAO
 
-    def parse_cap_map(s):
-        caps = {}
-        if not s: return caps
-        for item in s.split(","):
-            k,v = item.strip().split(":")
-            caps[k.strip()] = int(v)
-        return caps
-
     caps_map = parse_cap_map(args.cap_train_families)
 
     global pasta_origem, pasta_destino
@@ -341,12 +514,18 @@ def main():
     pasta_destino = os.path.join(modelo, "dataset", "split")
 
     soma = args.train + args.val + args.test
-    if soma <= 0: raise ValueError("Soma de proporções deve ser > 0.")
-    p_train = args.train / soma
-    p_val   = args.val   / soma
-    p_test  = args.test  / soma
+    if soma <= 0:
+        raise ValueError("Soma de proporções deve ser > 0.")
 
-    mins = {"train": max(0, args.min_train), "val": max(0, args.min_val), "test": max(0, args.min_test)}
+    p_train = args.train / soma
+    p_val = args.val / soma
+    p_test = args.test / soma
+
+    mins = {
+        "train": max(0, args.min_train),
+        "val": max(0, args.min_val),
+        "test": max(0, args.min_test),
+    }
 
     garantir(pasta_destino)
 
@@ -357,22 +536,28 @@ def main():
 
     random.seed(args.seed)
 
-    total_global = {"train":0, "val":0, "test":0, "familias":0}
+    total_global = {"train": 0, "val": 0, "test": 0, "familias": 0}
+
+    print(f"[INFO] Origem: {pasta_origem}")
+    print(f"[INFO] Destino: {pasta_destino}")
+    print(f"[INFO] dual_head_mask/masks2: {USE_MASKS2}")
+    print(f"[INFO] dual_head_label/labels: {USE_LABELS}")
     print(f"Grupos: {', '.join(grupos)}")
     print(f"Proporções normalizadas: train={p_train:.3f}, val={p_val:.3f}, test={p_test:.3f}")
-    print(f"Mínimos por grupo (famílias): train={mins['train']} val={mins['val']} test={mins['test']}")
+    print(f"Mínimos por grupo: train={mins['train']} val={mins['val']} test={mins['test']}")
 
     for g in grupos:
-        res = split_group(g, p_train, p_val, p_test, args.seed, mins, caps_map=caps_map)
+        res = split_group(g, p_train, p_val, p_test, args.seed, mins, caps_map=caps_map, strict_label=args.strict_label)
         for k in total_global.keys():
             total_global[k] += res.get(k, 0)
 
-    print("\nResumo global (imagens copiadas):")
+    print("\nResumo global, imagens copiadas:")
     print(f"  train: {total_global['train']}")
     print(f"  val:   {total_global['val']}")
     print(f"  test:  {total_global['test']}")
-    print(f"  famílias (total): {total_global['familias']}")
+    print(f"  famílias total: {total_global['familias']}")
     print("\n✅ Split sem vazamento concluído!")
 
+
 if __name__ == "__main__":
     main()

Python/OAK/datasets/_8_train_segformer.py

@@ -258,7 +258,7 @@ def run_one_epoch(model: nn.Module,
 
 def main():
     parser = argparse.ArgumentParser()
-    parser.add_argument("--config", default="config_oak.json")
+    parser.add_argument("--config", default="config.json")
     parser.add_argument("--epochs", type=int, default=120)
     parser.add_argument("--batch", type=int, default=1)  # <<< default seguro pra 8GB
     parser.add_argument("--lr", type=float, default=6e-5)

Python/OAK/datasets/_8_train_segformer_b3_dual.py

@@ -1,965 +0,0 @@
-# _8_train_segformer_b3_dual_v2.py
-# Treino SegFormer-B3 com 2 cabeças:
-#  - Head 1: segmentação semântica (3 classes + ignore)
-#  - Head 2: corredor binário (0/1 + ignore)
-#
-# Features:
-#  - class weights (opcional, estimado do train)
-#  - pos_weight (opcional, estimado do train p/ BCE do corredor)
-#  - warmup + cosine (opcional) + ReduceLROnPlateau
-#  - early stopping com métrica agregada
-#  - resume completo (model + corridor_head + optimizer + scaler)
-#
-# Observação: este script assume que seu ROISegDataset retorna dict com:
-#   {"image": <tensor/ndarray>, "mask": <tensor/ndarray>, ...}
-# e que você já acertou para também expor "image_path"/"mask_path" OU
-# que seu DualMaskROISegDataset consiga derivar paths corretamente.
-
-#python _8_train_segformer_b3_dual.py --epochs 80 --batch 2 --num_workers 2 --amp --amp_val --use_weights --auto_pos_weight --warmup_epochs 2 --cosine_epochs 15 --lambda_corr 0.5 --corr_dice_max 0.10 --corr_dice_ramp_epochs 10 --es_metric harmonic --es_patience 12
-
-#python _8_train_segformer_b3_dual.py --epochs 120 --batch 2 --num_workers 4 --amp --amp_val --grad_accum 1 --use_weights --auto_pos_weight --warmup_epochs 2 --cosine_epochs 20 --lambda_corr 0.7 --corr_dice_max 0.15 --corr_dice_ramp_epochs 10 --es_metric harmonic --es_patience 15
-
-import os
-import re
-import time
-import math
-import json
-import random
-import argparse
-from dataclasses import dataclass
-from typing import Any, Dict, List, Optional, Tuple
-
-import numpy as np
-from PIL import Image
-
-import torch
-import torch.nn as nn
-import torch.nn.functional as F
-from torch.utils.data import DataLoader
-
-from torch.amp import autocast
-from torch.cuda.amp import GradScaler
-
-# HuggingFace transformers
-from transformers import SegformerForSemanticSegmentation
-
-# Seu dataset (o mesmo que você já usa no projeto)
-from roi_seg_dataset import ROISegDataset
-
-
-# -----------------------------
-# Utils
-# -----------------------------
-def seed_everything(seed: int = 42):
-    random.seed(seed)
-    np.random.seed(seed)
-    torch.manual_seed(seed)
-    torch.cuda.manual_seed_all(seed)
-
-
-IMAGENET_MEAN = torch.tensor([0.485, 0.456, 0.406]).view(3, 1, 1)
-IMAGENET_STD  = torch.tensor([0.229, 0.224, 0.225]).view(3, 1, 1)
-
-
-def normalize_img(img: torch.Tensor) -> torch.Tensor:
-    return (img - IMAGENET_MEAN.to(img.device)) / IMAGENET_STD.to(img.device)
-
-
-@torch.no_grad()
-def update_confusion_matrix(cm: torch.Tensor, preds: torch.Tensor, labels: torch.Tensor, num_classes: int, ignore_index: int = 255):
-    preds = preds.view(-1)
-    labels = labels.view(-1)
-
-    valid = labels != ignore_index
-    preds = preds[valid]
-    labels = labels[valid]
-
-    idx = labels * num_classes + preds
-    bins = torch.bincount(idx, minlength=num_classes * num_classes)
-    cm += bins.view(num_classes, num_classes)
-
-
-@torch.no_grad()
-def compute_iou_from_cm(cm: torch.Tensor, eps: float = 1e-6) -> Tuple[float, List[float]]:
-    cm = cm.float()
-    tp = torch.diag(cm)
-    fp = cm.sum(0) - tp
-    fn = cm.sum(1) - tp
-    denom = tp + fp + fn + eps
-    iou = (tp / denom).cpu().tolist()
-    miou = float(np.mean(iou))
-    return miou, iou
-
-
-@torch.no_grad()
-def compute_pixel_acc_from_cm(cm: torch.Tensor, eps: float = 1e-6) -> float:
-    cm = cm.float()
-    acc = (torch.diag(cm).sum() / (cm.sum() + eps)).item()
-    return acc
-
-
-def default_collate_dual(batch):
-    imgs, m1s, m2s = [], [], []
-    for item in batch:
-        # item pode vir como (img, mask1, mask2) ou dict etc.
-        if isinstance(item, dict):
-            img = item["image"]
-            mask1 = item["mask1"]
-            mask2 = item["mask2"]
-        else:
-            img, mask1, mask2 = item
-
-        if isinstance(img, np.ndarray):
-            img = torch.from_numpy(img)
-        if isinstance(mask1, np.ndarray):
-            mask1 = torch.from_numpy(mask1)
-        if isinstance(mask2, np.ndarray):
-            mask2 = torch.from_numpy(mask2)
-
-        if img.ndim == 3 and img.shape[-1] == 3:
-            img = img.permute(2, 0, 1)
-
-        if img.dtype != torch.float32:
-            img = img.float()
-            if img.max() > 1.5:
-                img = img / 255.0
-
-        imgs.append(img)
-        m1s.append(mask1.long())
-        m2s.append(mask2.long())
-
-    return torch.stack(imgs, 0), torch.stack(m1s, 0), torch.stack(m2s, 0)
-
-
-def estimate_class_weights_from_ds(ds: ROISegDataset, num_classes: int, ignore_index: int = 255, max_samples: int = 800) -> torch.Tensor:
-    n = min(len(ds), max_samples)
-    idxs = np.random.choice(len(ds), size=n, replace=False)
-
-    counts = np.zeros(num_classes, dtype=np.float64)
-    for i in idxs:
-        item = ds[i]
-        mask = item["mask"] if isinstance(item, dict) else item[1]
-        m = mask.cpu().numpy() if isinstance(mask, torch.Tensor) else np.array(mask)
-
-        m = m.reshape(-1)
-        m = m[m != ignore_index]
-        if m.size == 0:
-            continue
-        counts += np.bincount(m, minlength=num_classes)[:num_classes]
-
-    freq = counts / (counts.sum() + 1e-12)
-    freq = np.clip(freq, 1e-12, 1.0)
-    weights = 1.0 / np.log(1.02 + freq)
-    weights = weights / weights.mean()
-    return torch.tensor(weights, dtype=torch.float32)
-
-
-def estimate_pos_weight_from_masks2(ds_dual, ignore_index2: int = 255, max_samples: int = 800) -> Optional[torch.Tensor]:
-    """
-    Estima pos_weight = neg/pos para BCEWithLogitsLoss.
-    Considera targets binários {0,1} ignorando ignore_index2.
-    """
-    n = min(len(ds_dual), max_samples)
-    if n <= 0:
-        return None
-
-    idxs = np.random.choice(len(ds_dual), size=n, replace=False)
-    pos = 0
-    neg = 0
-    for i in idxs:
-        it = ds_dual[i]
-        m2 = it["mask2"] if isinstance(it, dict) else it[2]
-
-        if isinstance(m2, torch.Tensor):
-            m2 = m2.cpu().numpy()
-        m2 = np.array(m2).reshape(-1)
-        m2 = m2[m2 != ignore_index2]
-        if m2.size == 0:
-            continue
-        pos += int((m2 == 1).sum())
-        neg += int((m2 == 0).sum())
-
-    if pos <= 0:
-        return None
-    pw = float(neg) / float(pos)
-    # clamp leve pra não explodir treino
-    pw = float(np.clip(pw, 1.0, 50.0))
-    return torch.tensor([pw], dtype=torch.float32)
-
-
-def dice_loss_with_logits(logits: torch.Tensor, targets: torch.Tensor, ignore_index: int = 255, eps: float = 1e-6) -> torch.Tensor:
-    """
-    logits: [B,1,H,W]
-    targets: [B,H,W] com {0,1} e possivelmente ignore_index
-    """
-    probs = torch.sigmoid(logits)
-    t = targets.float().unsqueeze(1)
-
-    if ignore_index is not None:
-        valid = (targets != ignore_index).unsqueeze(1)
-        probs = probs[valid]
-        t = t[valid]
-
-    probs = probs.reshape(-1)
-    t = t.reshape(-1)
-
-    inter = (probs * t).sum()
-    denom = probs.sum() + t.sum() + eps
-    dice = (2.0 * inter + eps) / denom
-    return 1.0 - dice
-
-
-@torch.no_grad()
-def binary_iou_and_acc_from_logits(logits: torch.Tensor, targets: torch.Tensor, thr: float = 0.5, ignore_index: int = 255, eps: float = 1e-6) -> Tuple[float, float]:
-    probs = torch.sigmoid(logits)
-    preds = (probs >= thr).long().squeeze(1)
-
-    if logits.shape[-2:] != targets.shape[-2:]:
-        preds = F.interpolate(preds.unsqueeze(1).float(), size=targets.shape[-2:], mode="nearest").long().squeeze(1)
-
-    valid = (targets != ignore_index)
-    if valid.sum().item() == 0:
-        return 0.0, 0.0
-
-    p = preds[valid]
-    t = targets[valid].long()
-
-    tp = ((p == 1) & (t == 1)).sum().float()
-    fp = ((p == 1) & (t == 0)).sum().float()
-    fn = ((p == 0) & (t == 1)).sum().float()
-
-    iou = (tp / (tp + fp + fn + eps)).item()
-    acc = ((p == t).sum().float() / (t.numel() + eps)).item()
-    return iou, acc
-
-
-def _pretty_iou(iou_list: List[float], class_name_by_id: Dict[int, str]) -> str:
-    parts = []
-    for cid, v in enumerate(iou_list):
-        parts.append(f"{class_name_by_id.get(cid, str(cid))}:{v:.3f}")
-    return " | ".join(parts)
-
-
-def save_checkpoint(path: str,
-                    base_model: nn.Module,
-                    corridor_head: nn.Module,
-                    optimizer: torch.optim.Optimizer,
-                    scaler: Optional[GradScaler],
-                    epoch: int,
-                    best_miou: float,
-                    best_main_iou: float,
-                    best_corr_iou: float,
-                    extra: Optional[Dict[str, Any]] = None):
-    ckpt = {
-        "epoch": epoch,
-        "model": base_model.state_dict(),
-        "corridor_head": corridor_head.state_dict(),
-        "optimizer": optimizer.state_dict(),
-        "best_miou": best_miou,
-        "best_main_iou": best_main_iou,
-        "best_corr_iou": best_corr_iou,
-    }
-    if scaler is not None:
-        ckpt["scaler"] = scaler.state_dict()
-    if extra:
-        ckpt["extra"] = extra
-    torch.save(ckpt, path)
-
-
-def load_checkpoint(path: str,
-                    base_model: nn.Module,
-                    corridor_head: nn.Module,
-                    optimizer: Optional[torch.optim.Optimizer] = None,
-                    scaler: Optional[GradScaler] = None,
-                    map_location: str = "cpu") -> Dict[str, Any]:
-    ckpt = torch.load(path, map_location=map_location, weights_only=False)
-    base_model.load_state_dict(ckpt["model"], strict=True)
-    corridor_head.load_state_dict(ckpt["corridor_head"], strict=True)
-    if optimizer is not None and "optimizer" in ckpt:
-        optimizer.load_state_dict(ckpt["optimizer"])
-    if scaler is not None and "scaler" in ckpt:
-        scaler.load_state_dict(ckpt["scaler"])
-    return ckpt
-
-
-# -----------------------------
-# Dual Dataset Wrapper
-# -----------------------------
-class DualMaskROISegDataset(torch.utils.data.Dataset):
-    def __init__(self, base_ds: ROISegDataset, split_root: str, ignore_index2: int = 255):
-        self.base_ds = base_ds
-        self.split_root = split_root
-        self.ignore_index2 = ignore_index2
-
-        # NÃO falha aqui por pasta. Vamos resolver por caminho da mask1.
-        self.masks2_dir = os.path.join(split_root, "masks2")
-        self.has_flat_masks2 = os.path.isdir(self.masks2_dir)
-
-    def __len__(self):
-        return len(self.base_ds)
-
-    def _get_mask1_path(self, item: Dict[str, Any], idx: int) -> str:
-        # 1) se vier no dict, usa
-        for k in ("mask_path", "mask_file", "maskname", "mask_name"):
-            if k in item and item[k]:
-                return str(item[k])
-
-        # 2) fallback: usa o índice no ROISegDataset (ele tem msk_paths)
-        if hasattr(self.base_ds, "msk_paths"):
-            return str(self.base_ds.msk_paths[idx])
-
-        raise RuntimeError(
-            "ROISegDataset não retornou mask_path/mask_file e não possui base_ds.msk_paths. "
-            "Não dá pra localizar a máscara para buscar a masks2."
-        )
-
-    def _load_mask2_from_mask1_path(self, mask1_path: str, target_shape_hw: Tuple[int, int]) -> torch.Tensor:
-        # 1) modo robusto: espelha estrutura trocando /masks/ -> /masks2/
-        norm = os.path.normpath(mask1_path)
-        parts = norm.split(os.sep)
-        try:
-            i = parts.index("masks")
-            parts[i] = "masks2"
-            path2 = os.sep.join(parts)
-        except ValueError:
-            path2 = ""
-
-        # 2) fallback: modo “flat” (split_root/masks2/<filename>)
-        if (not path2) or (not os.path.exists(path2)):
-            if not self.has_flat_masks2:
-                raise FileNotFoundError(
-                    f"Mask2 não encontrada espelhando a mask1.\n"
-                    f"mask1: {mask1_path}\n"
-                    f"tentativa: {path2}\n"
-                    f"E também não existe: {self.masks2_dir}"
-                )
-            filename = os.path.basename(mask1_path)
-            path2 = os.path.join(self.masks2_dir, filename)
-            if not os.path.exists(path2):
-                raise FileNotFoundError(f"Mask2 não encontrada: {path2}")
-
-        m = np.array(Image.open(path2))
-        if m.ndim == 3:
-            m = m[..., 0]
-
-        if m.max() > 1:
-            m = (m >= 128).astype(np.uint8)
-
-        H, W = target_shape_hw
-        if m.shape[0] != H or m.shape[1] != W:
-            m = np.array(Image.fromarray(m).resize((W, H), resample=Image.NEAREST))
-
-        return torch.from_numpy(m).long()
-
-    def __getitem__(self, idx):
-        item = self.base_ds[idx]
-        # caso venha (dict,) ou [dict]
-        if isinstance(item, (list, tuple)) and len(item) == 1 and isinstance(item[0], dict):
-            item = item[0]
-
-        if isinstance(item, dict):
-            pass
-        elif isinstance(item, (list, tuple)) and len(item) >= 2:
-            img, mask = item[0], item[1]
-            item = {"image": img, "mask": mask}
-        else:
-            raise RuntimeError(
-                f"ROISegDataset retornou formato inesperado: type={type(item)} value={item}"
-            )
-
-        img = item["image"]
-        mask1 = item["mask"]
-
-        if isinstance(mask1, torch.Tensor):
-            h, w = int(mask1.shape[-2]), int(mask1.shape[-1])
-        else:
-            m = np.array(mask1)
-            h, w = m.shape[0], m.shape[1]
-
-        mask1_path = self._get_mask1_path(item, idx)
-        mask2 = self._load_mask2_from_mask1_path(mask1_path, (h, w))
-
-        item["mask1"] = mask1
-        item["mask2"] = mask2
-        return item
-
-
-# -----------------------------
-# Corridor head
-# -----------------------------
-class CorridorHead(nn.Module):
-    """
-    Head simples:
-      feat (B,C,H,W) + logits_seg (B,K,H,W) -> concat -> convs -> 1 canal (logit corredor)
-    """
-    def __init__(self, feat_ch: int, num_classes: int, hidden: int = 256, dropout: float = 0.1):
-        super().__init__()
-        in_ch = feat_ch + num_classes
-        self.in_ch = in_ch
-        self.net = nn.Sequential(
-            nn.Conv2d(in_ch, hidden, kernel_size=3, padding=1),
-            nn.BatchNorm2d(hidden),
-            nn.ReLU(inplace=True),
-            nn.Dropout2d(dropout),
-            nn.Conv2d(hidden, hidden, kernel_size=3, padding=1),
-            nn.BatchNorm2d(hidden),
-            nn.ReLU(inplace=True),
-            nn.Dropout2d(dropout),
-            nn.Conv2d(hidden, 1, kernel_size=1)
-        )
-
-    def forward(self, feat: torch.Tensor, logits_seg: torch.Tensor) -> torch.Tensor:
-        x = torch.cat([feat, logits_seg], dim=1)
-        return self.net(x)
-
-
-# -----------------------------
-# One epoch
-# -----------------------------
-def run_one_epoch_dual(base_model: nn.Module,
-                      corridor_head: nn.Module,
-                      loader: DataLoader,
-                      optimizer: Optional[torch.optim.Optimizer],
-                      device: torch.device,
-                      num_classes: int,
-                      ignore_index1: int,
-                      ignore_index2: int,
-                      criterion_seg: nn.Module,
-                      bce_corr: nn.Module,
-                      lambda_corr: float,
-                      amp: bool,
-                      scaler: Optional[GradScaler],
-                      train: bool,
-                      grad_accum: int = 1,
-                      corridor_thr: float = 0.5,
-                      corr_dice_mix: float = 0.0) -> Dict[str, Any]:
-
-    base_model.train(train)
-    corridor_head.train(train)
-
-    total_loss = 0.0
-    total_loss_seg = 0.0
-    total_loss_corr = 0.0
-
-    cm = torch.zeros((num_classes, num_classes), dtype=torch.int64, device=device)
-
-    corr_iou_accum = 0.0
-    corr_acc_accum = 0.0
-
-    t0 = time.time()
-    n_batches = 0
-
-    with torch.set_grad_enabled(train):
-        if train and optimizer is not None:
-            optimizer.zero_grad(set_to_none=True)
-
-        for step, (imgs, masks1, masks2) in enumerate(loader, start=1):
-            imgs = imgs.to(device, non_blocking=True)
-            masks1 = masks1.to(device, non_blocking=True)
-            masks2 = masks2.to(device, non_blocking=True)
-
-            #print("masks2 unique:", torch.unique(masks2)[:10])
-            #print("masks2 max:", masks2.max().item(), "min:", masks2.min().item())
-
-            imgs = normalize_img(imgs)
-
-            with autocast(device_type="cuda", enabled=amp and device.type == "cuda"):
-                out = base_model(pixel_values=imgs)
-                logits1 = out.logits  # [B,K,H,W]
-
-                if logits1.shape[-2:] != masks1.shape[-2:]:
-                    logits1 = F.interpolate(logits1, size=masks1.shape[-2:], mode="bilinear", align_corners=False)
-
-                loss_seg = criterion_seg(logits1, masks1)
-
-                # feat do encoder: pega hidden_states (se disponível), senão reusa logits como proxy
-                feat = None
-                if hasattr(out, "hidden_states") and out.hidden_states is not None:
-                    feat = out.hidden_states[-1]  # [B, C, h, w]
-                else:
-                    # fallback: usa logits como feature (não ideal, mas mantém vivo)
-                    feat = logits1
-
-                if feat.shape[-2:] != logits1.shape[-2:]:
-                    feat = F.interpolate(feat, size=logits1.shape[-2:], mode="bilinear", align_corners=False)
-
-                logits_corr = corridor_head(feat, logits1)  # [B,1,H,W]
-
-                if logits_corr.shape[-2:] != masks2.shape[-2:]:
-                    logits_corr = F.interpolate(logits_corr, size=masks2.shape[-2:], mode="bilinear", align_corners=False)
-
-                #with torch.no_grad():
-                #    p = torch.sigmoid(logits_corr)
-                #    print("corr prob mean:", p.mean().item(), "min:", p.min().item(), "max:", p.max().item())
-
-                # BCE (com ignore via masking manual)
-                valid = (masks2 != ignore_index2)
-                if valid.sum().item() > 0:
-                    tgt = masks2.float().unsqueeze(1)
-                    bce_val = bce_corr(logits_corr[valid.unsqueeze(1)], tgt[valid.unsqueeze(1)])
-                    if corr_dice_mix > 0:
-                        d_val = dice_loss_with_logits(logits_corr, masks2, ignore_index=ignore_index2)
-                        loss_corr = (1.0 - corr_dice_mix) * bce_val + corr_dice_mix * d_val
-                    else:
-                        loss_corr = bce_val
-                else:
-                    loss_corr = torch.zeros([], device=device, dtype=loss_seg.dtype)
-
-                loss = loss_seg + lambda_corr * loss_corr
-
-                if train and grad_accum > 1:
-                    loss = loss / grad_accum
-
-            if train and optimizer is not None:
-                if amp and scaler is not None and device.type == "cuda":
-                    scaler.scale(loss).backward()
-                    if (step % grad_accum) == 0:
-                        scaler.step(optimizer)
-                        scaler.update()
-                        optimizer.zero_grad(set_to_none=True)
-                else:
-                    loss.backward()
-                    if (step % grad_accum) == 0:
-                        optimizer.step()
-                        optimizer.zero_grad(set_to_none=True)
-
-            total_loss += float(loss.item()) * (grad_accum if (train and grad_accum > 1) else 1.0)
-            total_loss_seg += float(loss_seg.item())
-            total_loss_corr += float(loss_corr.item())
-            n_batches += 1
-
-            # metrics seg
-            preds1 = torch.argmax(logits1.detach(), dim=1)
-            update_confusion_matrix(cm, preds1, masks1, num_classes, ignore_index=ignore_index1)
-
-            # metrics corr
-            ciou, cacc = binary_iou_and_acc_from_logits(logits_corr.detach(), masks2, thr=corridor_thr, ignore_index=ignore_index2)
-            corr_iou_accum += ciou
-            corr_acc_accum += cacc
-
-    miou, iou_per_class = compute_iou_from_cm(cm)
-    acc = compute_pixel_acc_from_cm(cm)
-
-    out = {
-        "loss": total_loss / max(1, n_batches),
-        "loss_seg": total_loss_seg / max(1, n_batches),
-        "loss_corr": total_loss_corr / max(1, n_batches),
-        "acc": acc,
-        "miou": miou,
-        "iou_per_class": iou_per_class,
-        "corr_iou": corr_iou_accum / max(1, n_batches),
-        "corr_acc": corr_acc_accum / max(1, n_batches),
-        "time_s": time.time() - t0
-    }
-    return out
-
-
-# -----------------------------
-# LR / EarlyStop helpers
-# -----------------------------
-@dataclass
-class EarlyStopState:
-    best: float = -1e9
-    bad_epochs: int = 0
-    stopped: bool = False
-
-
-def agg_metric(es_metric: str, miou: float, main_iou: float, corr_iou: float) -> float:
-    if es_metric == "miou":
-        return float(miou)
-    if es_metric == "main":
-        return float(main_iou)
-    if es_metric == "corr":
-        return float(corr_iou)
-    if es_metric == "harmonic":
-        eps = 1e-6
-        a = max(eps, float(main_iou))
-        b = max(eps, float(corr_iou))
-        return float(2.0 * a * b / (a + b + eps))
-    # default
-    return float(miou)
-
-
-def apply_warmup(optimizer, base_lr: float, epoch: int, warmup_epochs: int):
-    if warmup_epochs <= 0:
-        return
-    if epoch <= warmup_epochs:
-        wf = epoch / float(warmup_epochs)
-        for g in optimizer.param_groups:
-            g["lr"] = base_lr * wf
-
-
-# -----------------------------
-# Main
-# -----------------------------
-def main():
-    ap = argparse.ArgumentParser()
-
-    ap.add_argument("--config", default="config_oak.json")
-    ap.add_argument("--epochs", type=int, default=120)
-    ap.add_argument("--batch", type=int, default=4)
-    ap.add_argument("--num_workers", type=int, default=4)
-    ap.add_argument("--seed", type=int, default=42)
-
-    ap.add_argument("--lr", type=float, default=6e-5)
-    ap.add_argument("--wd", type=float, default=1e-2)
-    ap.add_argument("--min_lr", type=float, default=6e-6)
-
-    ap.add_argument("--amp", action="store_true")
-    ap.add_argument("--amp_val", action="store_true")
-    ap.add_argument("--grad_accum", type=int, default=1)
-
-    ap.add_argument("--ignore_index", type=int, default=255)
-    ap.add_argument("--ignore_index2", type=int, default=255)
-
-    ap.add_argument("--lambda_corr", type=float, default=0.5)
-    ap.add_argument("--corr_thr", type=float, default=0.5)
-
-    # class weights
-    ap.add_argument("--use_weights", action="store_true")
-    ap.add_argument("--cw_max_samples", type=int, default=800)
-
-    # pos_weight corredor
-    ap.add_argument("--auto_pos_weight", action="store_true")
-    ap.add_argument("--pw_max_samples", type=int, default=800)
-
-    # mistura dice no corredor (rampa)
-    ap.add_argument("--corr_dice_max", type=float, default=0.15)
-    ap.add_argument("--corr_dice_ramp_epochs", type=int, default=10)
-
-    # schedulers
-    ap.add_argument("--warmup_epochs", type=int, default=0)
-    ap.add_argument("--cosine_epochs", type=int, default=0, help="se >0: usa cosine por N épocas e depois troca p/ plateau")
-    ap.add_argument("--plateau_patience", type=int, default=3)
-    ap.add_argument("--plateau_factor", type=float, default=0.6)
-    ap.add_argument("--plateau_cooldown", type=int, default=1)
-
-    # early stopping
-    ap.add_argument("--es_metric", default="harmonic", choices=["miou", "main", "corr", "harmonic"])
-    ap.add_argument("--es_patience", type=int, default=12)
-    ap.add_argument("--es_min_delta", type=float, default=2e-4)
-
-    # logs/ckpt
-    ap.add_argument("--save_every", type=int, default=0)
-    ap.add_argument("--resume", action="store_true")
-
-    ap.add_argument("--main_class", type=str, default=None)
-
-    args = ap.parse_args()
-
-    seed_everything(args.seed)
-
-    with open(args.config, "r") as f:
-        config = json.load(f)
-
-    MODELO = config["camera"]
-    MODEL_NAME = config["model_name"]
-    RESOLUCAO = config["resolucao"]
-    ROI_INICIO = config["roi_inicio"]
-    ROI_TAMANHO = config["roi_tamanho"]
-    BACKBONE = config["backbone"]
-    MAIN_CLASS_NAME = str(config.get("main_class_name", "erva")).lower()
-    if args.main_class is not None:
-        MAIN_CLASS_NAME = args.main_class.lower()
-
-    dataset_path = os.path.join(MODELO, "dataset")
-    labelmap_path = os.path.join(dataset_path, "labelmap.txt")
-
-    # save paths
-    save_path = os.path.join(MODELO, "backup", config["modelo"], MODEL_NAME + "_dual")
-    os.makedirs(save_path, exist_ok=True)
-    last_ckpt_path = os.path.join(save_path, "last.pt")
-    best_miou_path = os.path.join(save_path, "best_miou.pt")
-    best_main_path = os.path.join(save_path, "best_main.pt")
-    best_corr_path = os.path.join(save_path, "best_corr.pt")
-
-    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-    print("Device:", device)
-
-    # seus splits estão em train/group/... então o ROISegDataset deve saber ler de lá
-    split_train = os.path.join(dataset_path, "split", "train")
-    split_val   = os.path.join(dataset_path, "split", "val")
-
-    ds_train_base = ROISegDataset(
-        split_train,
-        save_path, ROI_INICIO, ROI_TAMANHO,
-        RESOLUCAO[0], RESOLUCAO[1], labelmap_path
-    )
-
-    ds_val_base = ROISegDataset(
-        split_val,
-        save_path, ROI_INICIO, ROI_TAMANHO,
-        RESOLUCAO[0], RESOLUCAO[1], labelmap_path
-    )
-
-    ds_train = DualMaskROISegDataset(ds_train_base, split_train, ignore_index2=args.ignore_index2)
-    ds_val   = DualMaskROISegDataset(ds_val_base, split_val, ignore_index2=args.ignore_index2)
-
-    CLASS_NAMES = getattr(ds_train_base, "classes", None)
-    if CLASS_NAMES is None:
-        raise RuntimeError("ROISegDataset precisa expor .classes (list ou dict).")
-
-    if isinstance(CLASS_NAMES, list):
-        num_classes = len(CLASS_NAMES)
-        class_id_by_name = {n.lower(): i for i, n in enumerate(CLASS_NAMES)}
-        class_name_by_id = {i: n for i, n in enumerate(CLASS_NAMES)}
-    elif isinstance(CLASS_NAMES, dict):
-        if all(isinstance(k, int) for k in CLASS_NAMES.keys()):
-            num_classes = len(CLASS_NAMES)
-            class_name_by_id = {int(k): str(v) for k, v in CLASS_NAMES.items()}
-            class_id_by_name = {str(v).lower(): int(k) for k, v in CLASS_NAMES.items()}
-        else:
-            class_id_by_name = {str(k).lower(): int(v) for k, v in CLASS_NAMES.items()}
-            num_classes = len(class_id_by_name)
-            class_name_by_id = {v: k for k, v in class_id_by_name.items()}
-    else:
-        raise RuntimeError("Formato de .classes não reconhecido.")
-
-    main_class_id = class_id_by_name.get(MAIN_CLASS_NAME, None)
-    if main_class_id is None:
-        print(f"[WARN] main_class_name='{MAIN_CLASS_NAME}' não encontrado. best_main_iou usa mIoU.")
-    else:
-        print(f"Main class: '{MAIN_CLASS_NAME}' -> id={main_class_id}")
-
-
-    dl_train = DataLoader(
-        ds_train,
-        batch_size=args.batch,
-        shuffle=True,
-        num_workers=args.num_workers,
-        pin_memory=(device.type == "cuda"),
-        collate_fn=default_collate_dual
-    )
-    dl_val = DataLoader(
-        ds_val,
-        batch_size=args.batch,
-        shuffle=False,
-        num_workers=max(0, args.num_workers // 2),
-        pin_memory=(device.type == "cuda"),
-        collate_fn=default_collate_dual
-    )
-
-    # model base SegFormer
-    base_model = SegformerForSemanticSegmentation.from_pretrained(
-        BACKBONE,
-        num_labels=num_classes,
-        ignore_mismatched_sizes=True,
-        use_safetensors=True,  # <- evita torch.load do .bin
-    )
-    base_model.config.output_hidden_states = True
-    base_model.to(device)
-
-    # descobre feat_ch via dummy forward (mais robusto)
-    with torch.no_grad():
-        dummy = torch.zeros((1, 3, 512, 512), device=device)
-        out = base_model(pixel_values=dummy)
-        feat_ch = None
-        if hasattr(out, "hidden_states") and out.hidden_states is not None:
-            feat_ch = int(out.hidden_states[-1].shape[1])
-        else:
-            feat_ch = int(out.logits.shape[1])  # fallback
-    corridor_head = CorridorHead(feat_ch=feat_ch, num_classes=num_classes, hidden=256, dropout=0.1).to(device)
-    print(f"[corridor_head] in_ch = feat({feat_ch}) + logits_seg({num_classes}) = {corridor_head.in_ch}")
-
-    # losses
-    class_weights = None
-    if args.use_weights:
-        w = estimate_class_weights_from_ds(ds_train_base, num_classes=num_classes, ignore_index=args.ignore_index, max_samples=args.cw_max_samples)
-        class_weights = w.to(device)
-        print("[INFO] class_weights:", class_weights.detach().cpu().numpy().round(3).tolist())
-    criterion_seg = nn.CrossEntropyLoss(ignore_index=args.ignore_index, weight=class_weights)
-
-    pos_weight = None
-    if args.auto_pos_weight:
-        pw = estimate_pos_weight_from_masks2(ds_train, ignore_index2=args.ignore_index2, max_samples=args.pw_max_samples)
-        if pw is not None:
-            pos_weight = pw.to(device)
-            print("[INFO] pos_weight corredor (neg/pos):", float(pos_weight.item()))
-        else:
-            print("[INFO] pos_weight não estimável (sem positivos suficientes), usando None.")
-    bce_corr = nn.BCEWithLogitsLoss(pos_weight=pos_weight) if pos_weight is not None else nn.BCEWithLogitsLoss()
-
-    # optimizer (base + head)
-    params = list(base_model.parameters()) + list(corridor_head.parameters())
-    optimizer = torch.optim.AdamW(params, lr=args.lr, weight_decay=args.wd)
-
-    scaler = GradScaler(enabled=(args.amp and device.type == "cuda"))
-
-    # schedulers
-    cosine = None
-    if args.cosine_epochs and args.cosine_epochs > 0:
-        cosine = torch.optim.lr_scheduler.CosineAnnealingLR(
-            optimizer,
-            T_max=max(1, args.cosine_epochs),
-            eta_min=args.min_lr
-        )
-    plateau = torch.optim.lr_scheduler.ReduceLROnPlateau(
-        optimizer,
-        mode="min",
-        factor=args.plateau_factor,
-        patience=args.plateau_patience,
-        cooldown=args.plateau_cooldown,
-        min_lr=args.min_lr,
-        verbose=True
-    )
-    active_sched = "cosine" if cosine is not None else "plateau"
-
-    # resume
-    start_epoch = 1
-    best_miou = -1.0
-    best_main_iou = -1.0
-    best_corr_iou = -1.0
-
-    if args.resume and os.path.exists(last_ckpt_path):
-        ckpt = load_checkpoint(last_ckpt_path, base_model, corridor_head, optimizer=optimizer, scaler=scaler, map_location="cpu")
-        start_epoch = int(ckpt.get("epoch", 0)) + 1
-        best_miou = float(ckpt.get("best_miou", -1.0))
-        best_main_iou = float(ckpt.get("best_main_iou", -1.0))
-        best_corr_iou = float(ckpt.get("best_corr_iou", -1.0))
-        print(f"[RESUME] epoch={start_epoch} best_miou={best_miou:.4f} best_main_iou={best_main_iou:.4f} best_corr_iou={best_corr_iou:.4f}")
-
-    # early stopping
-    es = EarlyStopState(best=-1e9, bad_epochs=0, stopped=False)
-
-    for epoch in range(start_epoch, args.epochs + 1):
-        lr_now = optimizer.param_groups[0]["lr"]
-
-        # warmup
-        apply_warmup(optimizer, args.lr, epoch, args.warmup_epochs)
-        lr_now = optimizer.param_groups[0]["lr"]
-
-        # corr dice ramp
-        if args.corr_dice_max > 0 and args.corr_dice_ramp_epochs > 0:
-            corr_dice_mix = min(args.corr_dice_max, (epoch - 1) / float(args.corr_dice_ramp_epochs) * args.corr_dice_max)
-        else:
-            corr_dice_mix = 0.0
-
-        print(f"\n==== Epoch {epoch}/{args.epochs} | lr={lr_now:.2e} | lambda_corr={args.lambda_corr} | corr_dice={corr_dice_mix:.3f} | sched={active_sched} ====")
-
-        if device.type == "cuda":
-            torch.cuda.empty_cache()
-
-        tr = run_one_epoch_dual(
-            base_model=base_model,
-            corridor_head=corridor_head,
-            loader=dl_train,
-            optimizer=optimizer,
-            device=device,
-            num_classes=num_classes,
-            ignore_index1=args.ignore_index,
-            ignore_index2=args.ignore_index2,
-            criterion_seg=criterion_seg,
-            bce_corr=bce_corr,
-            lambda_corr=args.lambda_corr,
-            amp=args.amp,
-            scaler=scaler,
-            train=True,
-            grad_accum=max(1, args.grad_accum),
-            corridor_thr=args.corr_thr,
-            corr_dice_mix=corr_dice_mix
-        )
-
-        if device.type == "cuda":
-            torch.cuda.empty_cache()
-
-        va = run_one_epoch_dual(
-            base_model=base_model,
-            corridor_head=corridor_head,
-            loader=dl_val,
-            optimizer=None,
-            device=device,
-            num_classes=num_classes,
-            ignore_index1=args.ignore_index,
-            ignore_index2=args.ignore_index2,
-            criterion_seg=criterion_seg,
-            bce_corr=bce_corr,
-            lambda_corr=args.lambda_corr,
-            amp=args.amp_val,
-            scaler=None,
-            train=False,
-            grad_accum=1,
-            corridor_thr=args.corr_thr,
-            corr_dice_mix=corr_dice_mix
-        )
-
-        # scheduler step
-        if cosine is not None and epoch <= args.cosine_epochs:
-            cosine.step()
-        else:
-            active_sched = "plateau"
-            plateau.step(va["loss"])
-
-        main_iou = va["miou"] if main_class_id is None else va["iou_per_class"][main_class_id]
-        corr_iou = float(va["corr_iou"])
-
-        print(f"TRAIN: loss={tr['loss']:.4f} (seg={tr['loss_seg']:.4f} corr={tr['loss_corr']:.4f}) "
-              f"acc={tr['acc']:.4f} miou={tr['miou']:.4f} corr_iou={tr['corr_iou']:.4f} (t={tr['time_s']:.1f}s)")
-        print(f"VAL  : loss={va['loss']:.4f} (seg={va['loss_seg']:.4f} corr={va['loss_corr']:.4f}) "
-              f"acc={va['acc']:.4f} miou={va['miou']:.4f} main_iou={float(main_iou):.4f} "
-              f"corr_iou={va['corr_iou']:.4f} corr_acc={va['corr_acc']:.4f} (t={va['time_s']:.1f}s)")
-        print("IoU per class:", _pretty_iou(va["iou_per_class"], class_name_by_id))
-
-        # save last
-        save_checkpoint(
-            last_ckpt_path,
-            base_model,
-            corridor_head,
-            optimizer,
-            scaler=scaler,
-            epoch=epoch,
-            best_miou=best_miou,
-            best_main_iou=best_main_iou,
-            best_corr_iou=best_corr_iou,
-            extra={
-                "val_loss": va["loss"],
-                "val_miou": va["miou"],
-                "val_main_iou": float(main_iou),
-                "val_corr_iou": float(corr_iou),
-                "lr": optimizer.param_groups[0]["lr"],
-                "corr_dice_mix": corr_dice_mix
-            }
-        )
-
-        # save every
-        if args.save_every > 0 and (epoch % args.save_every == 0):
-            save_checkpoint(
-                os.path.join(save_path, f"epoch_{epoch:04d}.pt"),
-                base_model, corridor_head, optimizer, scaler=scaler,
-                epoch=epoch, best_miou=best_miou, best_main_iou=best_main_iou, best_corr_iou=best_corr_iou
-            )
-
-        # bests
-        if va["miou"] > best_miou:
-            best_miou = va["miou"]
-            save_checkpoint(best_miou_path, base_model, corridor_head, optimizer, scaler=scaler,
-                            epoch=epoch, best_miou=best_miou, best_main_iou=best_main_iou, best_corr_iou=best_corr_iou)
-            print(f"[BEST mIoU] {best_miou:.4f} -> saved: {best_miou_path}")
-
-        if float(main_iou) > best_main_iou:
-            best_main_iou = float(main_iou)
-            save_checkpoint(best_main_path, base_model, corridor_head, optimizer, scaler=scaler,
-                            epoch=epoch, best_miou=best_miou, best_main_iou=best_main_iou, best_corr_iou=best_corr_iou)
-            print(f"[BEST MAIN] {best_main_iou:.4f} -> saved: {best_main_path}")
-
-        if corr_iou > best_corr_iou:
-            best_corr_iou = corr_iou
-            save_checkpoint(best_corr_path, base_model, corridor_head, optimizer, scaler=scaler,
-                            epoch=epoch, best_miou=best_miou, best_main_iou=best_main_iou, best_corr_iou=best_corr_iou)
-            print(f"[BEST CORR] {best_corr_iou:.4f} -> saved: {best_corr_path}")
-
-        # early stopping
-        score = agg_metric(args.es_metric, va["miou"], float(main_iou), corr_iou)
-        if score > es.best + args.es_min_delta:
-            es.best = score
-            es.bad_epochs = 0
-            print(f"[ES] improved {args.es_metric} -> {score:.6f}")
-        else:
-            es.bad_epochs += 1
-            print(f"[ES] no improve ({es.bad_epochs}/{args.es_patience}) best={es.best:.6f} now={score:.6f}")
-            if es.bad_epochs >= args.es_patience:
-                print(f"🛑 Early stopping acionado (metric={args.es_metric}).")
-                break
-
-
-if __name__ == "__main__":
-    main()

Python/OAK/datasets/_9_test_segformer.py

@@ -215,7 +215,7 @@ def main():
     device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 
     # Lê config do projeto (mesmo padrão do fastscnn)
-    with open("config_oak.json", "r") as f:
+    with open("config.json", "r") as f:
         config = json.load(f)
 
     MODELO = config["camera"]
@@ -261,7 +261,7 @@ def main():
     model = load_segformer_from_checkpoint(pt_path, backbone=backbone, num_classes=num_classes, device=device)
 
     # Normalização (ImageNet, padrão de muita coisa; se teu treino usou outro, troca aqui)
-    from _8_train_segformer_b3 import normalize_img
+    from _8_train_segformer import normalize_img
 
     if args.camera:
         # === Modo câmera (igual estilo do fastscnn) ===

Python/OAK/datasets/_9_test_segformer_b3_dual.py

@@ -1,570 +0,0 @@
-import json
-import os
-import glob
-import argparse
-from dataclasses import dataclass
-from typing import List, Optional
-
-import cv2
-import numpy as np
-import torch
-import torch.nn as nn
-from transformers import SegformerForSemanticSegmentation
-
-
-# -----------------------------
-# Utils
-# -----------------------------
-def find_device():
-    return torch.device("cuda" if torch.cuda.is_available() else "cpu")
-
-def load_labelmap_lut_bgr(labelmap_path: str):
-    """
-    Lê labelmap no formato:
-      classe:r,g,b
-    Ignora linhas vazias e comentários '#'
-    Se existir 'ignore', retorna ignore_id=255 e ignore_bgr.
-
-    Retorna:
-      lut_bgr: dict[int, tuple(b,g,r)]
-      id_to_name: dict[int, str]
-      ignore_id: int (default 255)
-    """
-    lut_bgr = {}
-    id_to_name = {}
-    ignore_id = 255
-    ignore_bgr = (180, 0, 180)  # fallback
-
-    idx = 0
-    with open(labelmap_path, "r", encoding="utf-8") as f:
-        for line in f:
-            line = line.strip()
-            if not line or line.startswith("#"):
-                continue
-            parts = line.split(":")
-            if len(parts) < 2:
-                continue
-            name = parts[0].strip()
-            rgb_str = parts[1].strip()
-            r, g, b = map(int, rgb_str.split(","))
-            bgr = (b, g, r)
-
-            if name.lower() == "ignore":
-                ignore_bgr = bgr
-                continue
-
-            lut_bgr[idx] = bgr
-            id_to_name[idx] = name
-            idx += 1
-
-    lut_bgr[ignore_id] = ignore_bgr
-    return lut_bgr, id_to_name, ignore_id
-
-def imread_bgr(path: str) -> np.ndarray:
-    img = cv2.imread(path, cv2.IMREAD_COLOR)
-    if img is None:
-        raise RuntimeError(f"Falha ao ler imagem: {path}")
-    return img
-
-
-def imread_gray(path: str) -> np.ndarray:
-    m = cv2.imread(path, cv2.IMREAD_GRAYSCALE)
-    if m is None:
-        raise RuntimeError(f"Falha ao ler máscara: {path}")
-    return m
-
-
-def ensure_uint8(x: np.ndarray) -> np.ndarray:
-    if x.dtype == np.uint8:
-        return x
-    x = np.clip(x, 0, 255).astype(np.uint8)
-    return x
-
-
-def normalize01(x: np.ndarray, eps: float = 1e-6) -> np.ndarray:
-    x = x.astype(np.float32)
-    mn, mx = float(x.min()), float(x.max())
-    return (x - mn) / (mx - mn + eps)
-
-
-def colorize_seg(mask_ids: np.ndarray, lut_bgr: dict) -> np.ndarray:
-    h, w = mask_ids.shape[:2]
-    out = np.zeros((h, w, 3), dtype=np.uint8)
-
-    # pinta cada id presente
-    for cid, bgr in lut_bgr.items():
-        out[mask_ids == cid] = bgr
-
-    return out
-
-
-def overlay_mask(img_bgr: np.ndarray, mask_bgr: np.ndarray, alpha: float = 0.45) -> np.ndarray:
-    return cv2.addWeighted(img_bgr, 1.0 - alpha, mask_bgr, alpha, 0.0)
-
-
-def put_hud(img: np.ndarray, lines: List[str]) -> np.ndarray:
-    out = img.copy()
-    y = 22
-    for s in lines:
-        cv2.putText(out, s, (12, y), cv2.FONT_HERSHEY_SIMPLEX, 0.6, (15, 15, 15), 3, cv2.LINE_AA)
-        cv2.putText(out, s, (12, y), cv2.FONT_HERSHEY_SIMPLEX, 0.6, (245, 245, 245), 1, cv2.LINE_AA)
-        y += 22
-    return out
-
-
-# -----------------------------
-# Corridor Head (igual ao treino)
-# -----------------------------
-class CorridorHead(nn.Module):
-    """
-    Head simples:
-      feat (B,C,H,W) + logits_seg (B,K,H,W) -> concat -> convs -> 1 canal (logit corredor)
-    """
-    def __init__(self, feat_ch: int, num_classes: int, hidden: int = 256, dropout: float = 0.1):
-        super().__init__()
-        in_ch = feat_ch + num_classes
-        self.in_ch = in_ch
-        self.net = nn.Sequential(
-            nn.Conv2d(in_ch, hidden, kernel_size=3, padding=1),
-            nn.BatchNorm2d(hidden),
-            nn.ReLU(inplace=True),
-            nn.Dropout2d(dropout),
-            nn.Conv2d(hidden, hidden, kernel_size=3, padding=1),
-            nn.BatchNorm2d(hidden),
-            nn.ReLU(inplace=True),
-            nn.Dropout2d(dropout),
-            nn.Conv2d(hidden, 1, kernel_size=1)
-        )
-
-    def forward(self, feat: torch.Tensor, logits_seg: torch.Tensor) -> torch.Tensor:
-        x = torch.cat([feat, logits_seg], dim=1)
-        return self.net(x)
-
-
-# -----------------------------
-# Dataset discovery (split/test/group/**)
-# -----------------------------
-@dataclass
-class Sample:
-    img_path: str
-    mask_path: Optional[str]
-    mask2_path: Optional[str]
-    group_name: str
-    filename: str
-
-
-IMG_EXTS = (".png", ".jpg", ".jpeg", ".bmp", ".tif", ".tiff", ".webp")
-
-
-def discover_samples(split_root: str) -> List[Sample]:
-    """
-    Espera estrutura:
-      split/test/group/<qualquer>/images/*.png
-      split/test/group/<qualquer>/masks/*.png
-      split/test/group/<qualquer>/masks2/*.png   (opcional, mas recomendado)
-    """
-    group_root = os.path.join(split_root, "group")
-    if not os.path.isdir(group_root):
-        raise RuntimeError(f"Não achei pasta: {group_root}")
-
-    samples: List[Sample] = []
-
-    # pega qualquer images/ em qualquer subpasta de group
-    img_dirs = glob.glob(os.path.join(group_root, "**", "images"), recursive=True)
-    img_dirs = [d for d in img_dirs if os.path.isdir(d)]
-
-    for idir in img_dirs:
-        base = os.path.dirname(idir)  # .../group/<grupo_ou_subgrupo>
-        group_name = os.path.relpath(base, group_root).replace("\\", "/")
-
-        mdir = os.path.join(base, "masks")
-        m2dir = os.path.join(base, "masks2")
-
-        img_paths = []
-        for ext in IMG_EXTS:
-            img_paths.extend(glob.glob(os.path.join(idir, f"*{ext}")))
-        img_paths = sorted(img_paths)
-
-        def find_mask(dir_path: str, filename: str):
-            stem, _ = os.path.splitext(filename)
-            for ext in [".png", ".jpg", ".jpeg", ".bmp", ".tif"]:
-                p = os.path.join(dir_path, stem + ext)
-                if os.path.exists(p):
-                    return p
-            return None
-
-        for ip in img_paths:
-            fn = os.path.basename(ip)
-
-            mask_path  = find_mask(mdir, fn)
-            mask2_path = find_mask(m2dir, fn)
-
-            samples.append(Sample(
-                img_path=ip,
-                mask_path=mask_path,
-                mask2_path=mask2_path,
-                group_name=group_name,
-                filename=fn
-            ))
-
-    if len(samples) == 0:
-        raise RuntimeError(f"Nenhuma imagem encontrada em: {group_root}/**/images")
-    return samples
-
-
-# -----------------------------
-# Model loading
-# -----------------------------
-def build_models(num_classes: int, device: torch.device, backbone: str):
-    base_model = SegformerForSemanticSegmentation.from_pretrained(
-        backbone,
-        num_labels=num_classes,
-        ignore_mismatched_sizes=True,
-        use_safetensors=True,  # <- evita torch.load do .bin
-    )
-    base_model.to(device)
-    base_model.config.output_hidden_states = True
-
-    # descobre feat_ch real a partir de um forward fake
-    with torch.no_grad():
-        dummy = torch.zeros((1, 3, 512, 512), device=device)
-        out = base_model(pixel_values=dummy, output_hidden_states=True)
-        feat_ch = int(out.hidden_states[-1].shape[1])
-
-    in_ch = feat_ch + num_classes
-    corridor_head = CorridorHead(feat_ch=feat_ch, num_classes=num_classes, hidden=256, dropout=0.1)
-    print(f"[corridor_head] in_ch = feat({feat_ch}) + logits_seg({num_classes}) = {in_ch}")
-    return base_model, corridor_head
-
-
-def make_corridor_head(in_ch: int, mid: int = 256):
-    """
-    Cria CorridorHead independente de qual assinatura a classe tem no script.
-    Suporta:
-      - CorridorHead(in_ch=..., mid=...)
-      - CorridorHead(in_channels=..., hidden=...)
-      - CorridorHead(feat_ch=..., num_classes=...)  (quando in_ch = feat_ch + num_classes)
-    """
-    # 1) assinatura direta in_ch/mid
-    try:
-        return CorridorHead(in_ch=in_ch, mid=mid)
-    except TypeError:
-        pass
-
-    # 2) variações comuns
-    try:
-        return CorridorHead(in_channels=in_ch, mid=mid)
-    except TypeError:
-        pass
-
-    try:
-        return CorridorHead(in_channels=in_ch, hidden=mid)
-    except TypeError:
-        pass
-
-    # 3) assinatura "feat_ch + num_classes"
-    # aqui só funciona se for exatamente "feat_ch, num_classes"
-    # e se in_ch for decomponível (ex: 6 = 3+3 ou 515 = 512+3)
-    for feat_ch_guess, num_classes_guess in [(512, 3), (3, 3)]:
-        if feat_ch_guess + num_classes_guess == in_ch:
-            try:
-                return CorridorHead(feat_ch=feat_ch_guess, num_classes=num_classes_guess, hidden=mid)
-            except TypeError:
-                try:
-                    return CorridorHead(feat_ch=feat_ch_guess, num_classes=num_classes_guess)
-                except TypeError:
-                    pass
-
-    raise RuntimeError(
-        f"Não consegui instanciar CorridorHead para in_ch={in_ch}. "
-        f"Verifique a assinatura do __init__() da classe CorridorHead no script."
-    )
-
-def load_dual_checkpoint(ckpt_path, base_model, corridor_head, device):
-    ckpt = torch.load(ckpt_path, map_location="cpu", weights_only=False)
-
-    base_model.load_state_dict(ckpt["model"], strict=True)
-
-    # >>> pega quantos canais o head do checkpoint espera
-    sd = ckpt["corridor_head"]
-    in_ch_ckpt = int(sd["net.0.weight"].shape[1])  # ex: 6 ou 515
-
-    # >>> recria head se o in_ch não bater
-    in_ch_model = int(corridor_head.net[0].in_channels)
-    if in_ch_model != in_ch_ckpt:
-        print(f"[corridor_head] rebuild: in_ch_model={in_ch_model} -> in_ch_ckpt={in_ch_ckpt}")
-        corridor_head = make_corridor_head(in_ch=in_ch_ckpt, mid=256)
-
-    corridor_head.load_state_dict(sd, strict=True)
-
-    base_model.to(device).eval()
-    corridor_head.to(device).eval()
-    meta = ckpt.get("meta", {})
-    return meta, corridor_head
-
-
-# -----------------------------
-# Inference
-# -----------------------------
-def run_corridor_head(corridor_head, x_rgb, feat, logits_seg):
-    import inspect
-    import torch.nn.functional as F
-
-    n_args = len(inspect.signature(corridor_head.forward).parameters)
-
-    # descobre in_ch esperado pela primeira conv do head
-    in_ch = None
-    if hasattr(corridor_head, "net"):
-        try:
-            in_ch = int(corridor_head.net[0].in_channels)
-        except Exception:
-            in_ch = None
-
-    # prepara entradas conforme o head
-    if in_ch == 6:
-        # RGB(3) + probs(3)
-        probs = torch.softmax(logits_seg, dim=1)
-        if probs.shape[-2:] != x_rgb.shape[-2:]:
-            probs = F.interpolate(probs, size=x_rgb.shape[-2:], mode="bilinear", align_corners=False)
-        feat_in = x_rgb
-        logits_in = probs
-    else:
-        # feat(512) + logits(3)
-        if logits_seg.shape[-2:] != feat.shape[-2:]:
-            logits_in = F.interpolate(logits_seg, size=feat.shape[-2:], mode="bilinear", align_corners=False)
-        else:
-            logits_in = logits_seg
-        feat_in = feat
-
-    # chama do jeito que o head espera
-    if n_args == 2:
-        return corridor_head(feat_in, logits_in)
-    elif n_args == 1:
-        x2 = torch.cat([feat_in, logits_in], dim=1)
-        return corridor_head(x2)
-
-    raise RuntimeError(f"Assinatura inesperada: {inspect.signature(corridor_head.forward)}")
-
-@torch.no_grad()
-def infer_dual(base_model, corridor_head, img_bgr: np.ndarray, device: torch.device,
-              input_size: int = 512, corridor_thr: float = 0.5):
-    """
-    Retorna:
-      pred_ids_full: [H,W] uint8
-      prob_corr_full: [H,W] float32 (0..1)
-      bin_corr_full: [H,W] uint8 (0/255)
-    """
-    h0, w0 = img_bgr.shape[:2]
-
-    img_rgb = cv2.cvtColor(img_bgr, cv2.COLOR_BGR2RGB)
-    img_res = cv2.resize(img_rgb, (input_size, input_size), interpolation=cv2.INTER_AREA)
-
-    x = torch.from_numpy(img_res).float() / 255.0
-    x = x.permute(2, 0, 1).unsqueeze(0).to(device)
-
-    out = base_model(pixel_values=x, output_hidden_states=True)
-    logits_seg = out.logits              # [B, 3, H, W]
-    feat = out.hidden_states[-1]         # [B, 512, h, w] (não vamos usar no modo 6)
-    
-    logits_corr = run_corridor_head(corridor_head, x, feat, logits_seg)
-
-    prob_corr = torch.sigmoid(logits_corr)[0, 0].detach().float().cpu().numpy()
-    bin_corr = (prob_corr >= corridor_thr).astype(np.uint8) * 255
-
-    pred_ids = torch.argmax(logits_seg, dim=1)[0].detach().cpu().numpy().astype(np.uint8)
-
-    # volta pro tamanho original
-    pred_ids_full = cv2.resize(pred_ids, (w0, h0), interpolation=cv2.INTER_NEAREST)
-    prob_corr_full = cv2.resize(prob_corr, (w0, h0), interpolation=cv2.INTER_LINEAR)
-    bin_corr_full = cv2.resize(bin_corr, (w0, h0), interpolation=cv2.INTER_NEAREST)
-
-    return pred_ids_full, prob_corr_full, bin_corr_full
-
-
-# -----------------------------
-# Viewer
-# -----------------------------
-def make_panel(sample: Sample , lut_bgr: dict, img_bgr: np.ndarray,
-               gt1: Optional[np.ndarray],
-               pred1: np.ndarray,
-               gt2: Optional[np.ndarray],
-               prob2: np.ndarray,
-               bin2: np.ndarray,
-               corridor_thr: float):
-    h, w = img_bgr.shape[:2]
-
-    pred1_col = colorize_seg(pred1, lut_bgr)
-    pred1_ov = overlay_mask(img_bgr, pred1_col, 0.45)
-
-    if gt1 is not None:
-        gt1_col = colorize_seg(gt1, lut_bgr)
-        gt1_ov = overlay_mask(img_bgr, gt1_col, 0.45)
-    else:
-        gt1_ov = img_bgr.copy()
-        gt1_col = np.zeros_like(img_bgr)
-
-    # --- PRED CORR (prob) sólido: verde=navegável, vermelho=bloqueado
-    prob = np.clip(prob2, 0.0, 1.0).astype(np.float32)
-    g = (prob * 255.0).astype(np.uint8)
-    r = ((1.0 - prob) * 255.0).astype(np.uint8)
-    b = np.zeros_like(g, dtype=np.uint8)
-    prob_solid = cv2.merge([b, g, r])  # BGR
-
-    # (opcional) desenha contorno do binário por cima do prob
-    edges = cv2.Canny(bin2, 50, 150)
-    prob_solid[edges > 0] = (255, 255, 255)
-
-    # --- PRED CORR (bin) sólido: branco=navegável, preto=bloqueado
-    bin2_solid = cv2.cvtColor(bin2, cv2.COLOR_GRAY2BGR)
-
-    if gt2 is not None:
-        gt2_vis = (gt2.copy()).astype(np.uint8)
-        gt2_vis[gt2_vis == 1] = 255
-        gt2_bgr = cv2.cvtColor(gt2_vis, cv2.COLOR_GRAY2BGR)
-        gt2_ov = overlay_mask(img_bgr, gt2_bgr, 0.35)
-    else:
-        gt2_ov = img_bgr.copy()
-
-    # monta grid
-    tile_w = 520
-    tile_h = int(tile_w * h / w)
-
-    def fit(im):
-        return cv2.resize(im, (tile_w, tile_h), interpolation=cv2.INTER_AREA)
-
-    row1 = np.concatenate([fit(img_bgr), fit(gt1_ov), fit(pred1_ov)], axis=1)
-    row2 = np.concatenate([fit(gt2_ov), fit(prob_solid), fit(bin2_solid)], axis=1)
-    panel = np.concatenate([row1, row2], axis=0)
-
-    hud = [
-        f"{sample.group_name}/{sample.filename}",
-        f"Corr thr={corridor_thr:.2f} | A/D navega | Q sai",
-        "Topo: IMG | GT SEG | PRED SEG",
-        "Baixo: GT CORR | PRED CORR (prob) | PRED CORR (bin)"
-    ]
-    panel = put_hud(panel, hud)
-    return panel
-
-
-def main():
-    ap = argparse.ArgumentParser()
-    ap.add_argument("--ckpt", type=str, default=None, help="Caminho do .pt (se não informar, usa best_miou.pt)")
-    ap.add_argument("--split", default="test", choices=["test", "val"], help="qual split usar")
-    ap.add_argument("--input_size", type=int, default=512)
-    ap.add_argument("--corr_thr", type=float, default=0.5)
-    args = ap.parse_args()
-
-    device = find_device()
-    print("Device:", device)
-
-    with open("config_oak.json", "r") as f:
-        config = json.load(f)
-    MODELO = config["camera"]
-    MODEL_NAME    = config["model_name"]     # ex: "segformer_b3"
-    BACKBONE      = config["backbone"]
-    modelo_folder = config["modelo"]         # ex: "dif"
-    dataset_path = os.path.join(MODELO, "dataset")
-
-    split_root = os.path.join(dataset_path, "split", args.split)
-    if not os.path.isdir(split_root):
-        print(f"[WARN] split/{args.split} não existe. Vou usar split/val.")
-        split_root = os.path.join(dataset_path, "split", "val")
-
-    samples = discover_samples(split_root)
-    print(f"[INFO] samples: {len(samples)}  | split_root={split_root}")
-
-    labelmap_path = os.path.join(dataset_path, "labelmap.txt")  # ajusta se seu arquivo estiver em outro lugar
-    lut_bgr, id_to_name, ignore_id = load_labelmap_lut_bgr(labelmap_path)
-    print("[labelmap]", id_to_name)
-
-    # ---- Descobre checkpoint ----
-    if args.ckpt is not None:
-        ckpt_path = args.ckpt
-    else:
-        # Caminho padrão igual ao treino:
-        # save_path = MODELO/backup/modelo/model_name/raw4
-        save_path = os.path.join(MODELO, "backup", modelo_folder, f"{MODEL_NAME}_dual")
-        ckpt_path = os.path.join(save_path, "best_miou.pt")
-
-    if not os.path.isfile(ckpt_path):
-        raise SystemExit(f"Checkpoint não encontrado em: {ckpt_path}")
-
-    print(f"[model] ckpt = {ckpt_path}")
-
-    num_classes = len(id_to_name)
-    base_model, corridor_head = build_models(num_classes=num_classes, device=device, backbone=BACKBONE)
-    meta, corridor_head = load_dual_checkpoint(ckpt_path, base_model, corridor_head, device=device)
-    print(f"... epoch={meta.get('epoch')} best_miou={meta.get('best_miou')} best_main={meta.get('best_main_iou')} best_corr={meta.get('best_corr_iou')}")
-
-    idx = 0
-    win = "SegFormer Dual Viewer"
-    cv2.namedWindow(win, cv2.WINDOW_NORMAL)
-
-    while True:
-        s = samples[idx]
-        img = imread_bgr(s.img_path)
-
-        gt1 = imread_gray(s.mask_path) if s.mask_path else None
-        gt2 = imread_gray(s.mask2_path) if s.mask2_path else None
-
-        pred1, prob2, bin2 = infer_dual(
-            base_model=base_model,
-            corridor_head=corridor_head,
-            img_bgr=img,
-            device=device,
-            input_size=args.input_size,
-            corridor_thr=args.corr_thr
-        )
-
-        def dbg_mask(name, m):
-            if m is None:
-                print(f"[{name}] None")
-            else:
-                u = np.unique(m)
-                print(f"[{name}] shape={m.shape} dtype={m.dtype} unique={u[:20]}{'...' if len(u)>20 else ''}")
-
-        #dbg_mask("gt1(seg)", gt1)
-        #dbg_mask("gt2(corr)", gt2)
-
-        panel = make_panel(
-            sample=s,
-            lut_bgr=lut_bgr,
-            img_bgr=img,
-            gt1=gt1,
-            pred1=pred1,
-            gt2=gt2,
-            prob2=prob2,
-            bin2=bin2,
-            corridor_thr=args.corr_thr
-        )
-
-        cv2.imshow(win, panel)
-        k = cv2.waitKey(0) & 0xFF
-
-        # Q / ESC
-        if k in (ord('q'), 27):
-            break
-
-        # A (volta)
-        if k in (ord('a'), ord('A')):
-            idx = (idx - 1) % len(samples)
-            continue
-
-        # D (avança)
-        if k in (ord('d'), ord('D')):
-            idx = (idx + 1) % len(samples)
-            continue
-
-        # W/S muda threshold
-        if k in (ord('w'), ord('W')):
-            args.corr_thr = min(0.95, args.corr_thr + 0.05)
-            continue
-        if k in (ord('s'), ord('S')):
-            args.corr_thr = max(0.05, args.corr_thr - 0.05)
-            continue
-
-    cv2.destroyAllWindows()
-
-
-if __name__ == "__main__":
-    main()

Python/OAK/datasets/_9_test_segformer_dual.py

@@ -0,0 +1,998 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+Viewer/teste SegFormer com suporte a:
+
+1) single
+   - segmentação semântica
+
+2) dual_head_mask
+   - segmentação + mask2 binária
+
+3) dual_head_label
+   - segmentação + estado global do corredor
+
+Compatível com checkpoint salvo pelo script:
+  _8_train_segformer_dual_mask_or_label.py
+
+Uso:
+  python _9_test_segformer_dual_mask_or_label.py --config config.json --split test
+  python _9_test_segformer_dual_mask_or_label.py --config config.json --split val --ckpt caminho.pt
+  python _9_test_segformer_dual_mask_or_label.py --config config.json --split test --save_dir outputs_test
+
+Teclas no viewer:
+  D / seta direita = próxima
+  A / seta esquerda = anterior
+  W/S = muda threshold da mask2, apenas modo dual_head_mask
+  Q / ESC = sair
+"""
+
+import os
+import re
+import cv2
+import glob
+import json
+import argparse
+from dataclasses import dataclass
+from typing import Dict, List, Optional, Tuple, Any
+import time
+import depthai as dai
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from transformers import SegformerForSemanticSegmentation
+
+
+# ============================================================
+# Utils básicos
+# ============================================================
+
+IMAGENET_MEAN = torch.tensor([0.485, 0.456, 0.406]).view(3, 1, 1)
+IMAGENET_STD = torch.tensor([0.229, 0.224, 0.225]).view(3, 1, 1)
+def normalize_img(img: torch.Tensor) -> torch.Tensor:
+    return (img - IMAGENET_MEAN.to(img.device)) / IMAGENET_STD.to(img.device)
+
+
+def find_device():
+    return torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+
+def load_labelmap_lut_bgr(labelmap_path: str):
+    lut_bgr = {}
+    id_to_name = {}
+    ignore_id = 255
+    ignore_bgr = (180, 0, 180)
+    idx = 0
+
+    with open(labelmap_path, "r", encoding="utf-8") as f:
+        for line in f:
+            line = line.strip()
+            if not line or line.startswith("#"):
+                continue
+            parts = line.split(":")
+            if len(parts) < 2:
+                continue
+            name = parts[0].strip()
+            rgb_str = parts[1].strip()
+            r, g, b = map(int, rgb_str.split(","))
+            bgr = (b, g, r)
+
+            if name.lower() == "ignore":
+                ignore_bgr = bgr
+                continue
+
+            lut_bgr[idx] = bgr
+            id_to_name[idx] = name
+            idx += 1
+
+    lut_bgr[ignore_id] = ignore_bgr
+    return lut_bgr, id_to_name, ignore_id
+
+
+def imread_bgr(path: str) -> np.ndarray:
+    img = cv2.imread(path, cv2.IMREAD_COLOR)
+    if img is None:
+        raise RuntimeError(f"Falha ao ler imagem: {path}")
+    return img
+
+
+def imread_gray(path: str) -> np.ndarray:
+    m = cv2.imread(path, cv2.IMREAD_GRAYSCALE)
+    if m is None:
+        raise RuntimeError(f"Falha ao ler máscara: {path}")
+    return m
+
+
+def colorize_seg(mask_ids: np.ndarray, lut_bgr: dict) -> np.ndarray:
+    h, w = mask_ids.shape[:2]
+    out = np.zeros((h, w, 3), dtype=np.uint8)
+    for cid, bgr in lut_bgr.items():
+        out[mask_ids == cid] = bgr
+    return out
+
+
+def overlay_mask(img_bgr: np.ndarray, mask_bgr: np.ndarray, alpha: float = 0.45) -> np.ndarray:
+    return cv2.addWeighted(img_bgr, 1.0 - alpha, mask_bgr, alpha, 0.0)
+
+
+def put_hud(img: np.ndarray, lines: List[str]) -> np.ndarray:
+    out = img.copy()
+    y = 24
+    for s in lines:
+        cv2.putText(out, s, (12, y), cv2.FONT_HERSHEY_SIMPLEX, 0.62, (15, 15, 15), 3, cv2.LINE_AA)
+        cv2.putText(out, s, (12, y), cv2.FONT_HERSHEY_SIMPLEX, 0.62, (245, 245, 245), 1, cv2.LINE_AA)
+        y += 24
+    return out
+
+
+def resize_fit(im: np.ndarray, tile_w: int, tile_h: Optional[int] = None) -> np.ndarray:
+    h, w = im.shape[:2]
+    if tile_h is None:
+        tile_h = int(tile_w * h / max(1, w))
+    return cv2.resize(im, (tile_w, tile_h), interpolation=cv2.INTER_AREA)
+
+
+def safe_float(v, default=None):
+    try:
+        return float(v)
+    except Exception:
+        return default
+
+
+# ============================================================
+# Métricas
+# ============================================================
+
+def seg_metrics(pred: np.ndarray, gt: Optional[np.ndarray], num_classes: int, ignore_id: int = 255) -> Optional[Dict[str, Any]]:
+    if gt is None:
+        return None
+    if pred.shape != gt.shape:
+        pred = cv2.resize(pred, (gt.shape[1], gt.shape[0]), interpolation=cv2.INTER_NEAREST)
+
+    valid = gt != ignore_id
+    if valid.sum() == 0:
+        return None
+
+    ious = []
+    for c in range(num_classes):
+        p = (pred == c) & valid
+        t = (gt == c) & valid
+        inter = np.logical_and(p, t).sum()
+        union = np.logical_or(p, t).sum()
+        ious.append(float(inter / union) if union > 0 else 0.0)
+
+    acc = float((pred[valid] == gt[valid]).sum() / max(1, valid.sum()))
+    return {"miou": float(np.mean(ious)), "acc": acc, "ious": ious}
+
+
+def mask2_metrics(bin_pred_255: np.ndarray, gt: Optional[np.ndarray]) -> Optional[Dict[str, float]]:
+    if gt is None:
+        return None
+    if bin_pred_255.shape != gt.shape:
+        bin_pred_255 = cv2.resize(bin_pred_255, (gt.shape[1], gt.shape[0]), interpolation=cv2.INTER_NEAREST)
+
+    p = bin_pred_255 >= 128
+    t = gt >= 128
+    tp = np.logical_and(p, t).sum()
+    fp = np.logical_and(p, ~t).sum()
+    fn = np.logical_and(~p, t).sum()
+    union = tp + fp + fn
+    iou = float(tp / union) if union > 0 else 0.0
+    acc = float((p == t).sum() / p.size)
+    return {"iou": iou, "acc": acc}
+
+
+# ============================================================
+# Heads iguais ao treino novo
+# ============================================================
+
+class Mask2Head(nn.Module):
+    def __init__(self, feat_ch: int, num_classes: int, hidden: int = 256, dropout: float = 0.1):
+        super().__init__()
+        in_ch = feat_ch + num_classes
+        self.in_ch = in_ch
+        self.net = nn.Sequential(
+            nn.Conv2d(in_ch, hidden, kernel_size=3, padding=1),
+            nn.BatchNorm2d(hidden),
+            nn.ReLU(inplace=True),
+            nn.Dropout2d(dropout),
+            nn.Conv2d(hidden, hidden, kernel_size=3, padding=1),
+            nn.BatchNorm2d(hidden),
+            nn.ReLU(inplace=True),
+            nn.Dropout2d(dropout),
+            nn.Conv2d(hidden, 1, kernel_size=1),
+        )
+
+    def forward(self, feat: torch.Tensor, logits_seg: torch.Tensor) -> torch.Tensor:
+        x = torch.cat([feat, logits_seg], dim=1)
+        return self.net(x)
+
+
+class LabelHead(nn.Module):
+    def __init__(self, feat_ch: int, num_seg_classes: int, num_label_classes: int, hidden: int = 256, dropout: float = 0.2):
+        super().__init__()
+        in_ch = feat_ch + num_seg_classes
+        self.in_ch = in_ch
+        self.pool = nn.AdaptiveAvgPool2d((1, 1))
+        self.net = nn.Sequential(
+            nn.Linear(in_ch, hidden),
+            nn.ReLU(inplace=True),
+            nn.Dropout(dropout),
+            nn.Linear(hidden, num_label_classes),
+        )
+
+    def forward(self, feat: torch.Tensor, logits_seg: torch.Tensor) -> torch.Tensor:
+        if feat.shape[-2:] != logits_seg.shape[-2:]:
+            feat = F.interpolate(feat, size=logits_seg.shape[-2:], mode="bilinear", align_corners=False)
+        x = torch.cat([feat, logits_seg], dim=1)
+        x = self.pool(x).flatten(1)
+        return self.net(x)
+
+
+# ============================================================
+# Discovery
+# ============================================================
+
+@dataclass
+class Sample:
+    img_path: str
+    mask_path: Optional[str]
+    mask2_path: Optional[str]
+    label_json_path: Optional[str]
+    label_npy_path: Optional[str]
+    group_name: str
+    filename: str
+
+
+IMG_EXTS = (".png", ".jpg", ".jpeg", ".bmp", ".tif", ".tiff", ".webp")
+
+
+def find_by_stem(dir_path: str, filename: str, exts: Tuple[str, ...]) -> Optional[str]:
+    if not dir_path or not os.path.isdir(dir_path):
+        return None
+    stem, _ = os.path.splitext(filename)
+    for ext in exts:
+        p = os.path.join(dir_path, stem + ext)
+        if os.path.exists(p):
+            return p
+    return None
+
+
+def discover_samples(split_root: str) -> List[Sample]:
+    group_root = os.path.join(split_root, "group")
+    if not os.path.isdir(group_root):
+        raise RuntimeError(f"Não achei pasta: {group_root}")
+
+    samples: List[Sample] = []
+    img_dirs = glob.glob(os.path.join(group_root, "**", "images"), recursive=True)
+    img_dirs = [d for d in img_dirs if os.path.isdir(d)]
+
+    for idir in img_dirs:
+        base = os.path.dirname(idir)
+        group_name = os.path.relpath(base, group_root).replace("\\", "/")
+        mdir = os.path.join(base, "masks")
+        m2dir = os.path.join(base, "masks2")
+        ldir = os.path.join(base, "labels")
+
+        img_paths = []
+        for ext in IMG_EXTS:
+            img_paths.extend(glob.glob(os.path.join(idir, f"*{ext}")))
+        img_paths = sorted(img_paths)
+
+        for ip in img_paths:
+            fn = os.path.basename(ip)
+            mask_path = find_by_stem(mdir, fn, (".png", ".jpg", ".jpeg", ".bmp", ".tif"))
+            mask2_path = find_by_stem(m2dir, fn, (".png", ".jpg", ".jpeg", ".bmp", ".tif"))
+            label_npy_path = find_by_stem(ldir, fn, (".npy",))
+            label_json_path = find_by_stem(ldir, fn, (".json", ".txt"))
+
+            samples.append(Sample(
+                img_path=ip,
+                mask_path=mask_path,
+                mask2_path=mask2_path,
+                label_json_path=label_json_path,
+                label_npy_path=label_npy_path,
+                group_name=group_name,
+                filename=fn,
+            ))
+
+    if len(samples) == 0:
+        raise RuntimeError(f"Nenhuma imagem encontrada em: {group_root}/**/images")
+    return samples
+
+
+def discover_image_folder(folder: str) -> List[Sample]:
+    samples = []
+
+    img_paths = []
+    for ext in IMG_EXTS:
+        img_paths.extend(glob.glob(os.path.join(folder, f"*{ext}")))
+
+    for ip in sorted(img_paths):
+        samples.append(Sample(
+            img_path=ip,
+            mask_path=None,
+            mask2_path=None,
+            label_json_path=None,
+            label_npy_path=None,
+            group_name="external",
+            filename=os.path.basename(ip),
+        ))
+
+    if not samples:
+        raise RuntimeError(f"Nenhuma imagem encontrada em: {folder}")
+
+    return samples
+
+
+def read_gt_label(sample: Sample) -> Tuple[Optional[int], Optional[str]]:
+    if sample.label_npy_path and os.path.exists(sample.label_npy_path):
+        try:
+            v = np.load(sample.label_npy_path)
+            return int(np.array(v).reshape(-1)[0]), None
+        except Exception:
+            pass
+
+    if sample.label_json_path and os.path.exists(sample.label_json_path):
+        ext = os.path.splitext(sample.label_json_path)[1].lower()
+        if ext == ".json":
+            with open(sample.label_json_path, "r", encoding="utf-8") as f:
+                data = json.load(f)
+            lid = data.get("label_id")
+            lname = data.get("estado_corredor") or data.get("label") or data.get("state")
+            return (int(lid) if lid is not None else None), lname
+        if ext == ".txt":
+            with open(sample.label_json_path, "r", encoding="utf-8") as f:
+                txt = f.read().strip()
+            try:
+                return int(txt), None
+            except Exception:
+                return None, txt
+
+    return None, None
+
+
+# ============================================================
+# Model loading
+# ============================================================
+
+def build_base_model(num_classes: int, device: torch.device, backbone: str):
+    base_model = SegformerForSemanticSegmentation.from_pretrained(
+        backbone,
+        num_labels=num_classes,
+        ignore_mismatched_sizes=True,
+        use_safetensors=True,
+    )
+    base_model.config.output_hidden_states = True
+    base_model.to(device).eval()
+    return base_model
+
+
+def get_feat_ch(base_model, device, input_h=512, input_w=512) -> int:
+    with torch.no_grad():
+        dummy = torch.zeros((1, 3, input_h, input_w), device=device)
+        out = base_model(pixel_values=dummy, output_hidden_states=True)
+        return int(out.hidden_states[-1].shape[1])
+
+
+def infer_mode_from_config(config: dict) -> str:
+    use_mask = bool(config.get("dual_head_mask", config.get("dual_head", False)))
+    use_label = bool(config.get("dual_head_label", False))
+    if use_mask and use_label:
+        raise RuntimeError("Config com dual_head_mask e dual_head_label ativos ao mesmo tempo. Este viewer espera apenas um.")
+    if use_mask:
+        return "mask2"
+    if use_label:
+        return "label"
+    return "single"
+
+
+def default_save_path(config: dict, mode: str) -> str:
+    modelo = config["camera"]
+    model_name = config["model_name"]
+    modelo_folder = config["modelo"]
+    suffix = {
+        "single": "_single",
+        "mask2": "_dual_mask",
+        "label": "_dual_label",
+    }[mode]
+    return os.path.join(modelo, "backup", modelo_folder, model_name + suffix)
+
+
+def load_checkpoint(ckpt_path: str, base_model, aux_head, device):
+    ckpt = torch.load(ckpt_path, map_location="cpu", weights_only=False)
+
+    base_model.load_state_dict(ckpt["model"], strict=True)
+
+    if aux_head is not None:
+        if "aux_head" not in ckpt:
+            raise RuntimeError("Checkpoint não possui aux_head. Ele é compatível com single head apenas?")
+        aux_head.load_state_dict(ckpt["aux_head"], strict=True)
+
+    base_model.to(device).eval()
+    if aux_head is not None:
+        aux_head.to(device).eval()
+
+    return ckpt
+
+
+# ============================================================
+# Inference
+# ============================================================
+
+def preprocess_img(img_bgr: np.ndarray, input_size: int, device: torch.device):
+    img_rgb = cv2.cvtColor(img_bgr, cv2.COLOR_BGR2RGB)
+    img_res = cv2.resize(img_rgb, (input_size, input_size), interpolation=cv2.INTER_AREA)
+    x = torch.from_numpy(img_res).float() / 255.0
+    x = x.permute(2, 0, 1).unsqueeze(0).to(device)
+    img_tensor = normalize_img(x)
+    return img_tensor
+
+
+@torch.no_grad()
+def infer_sample(base_model, aux_head, img_bgr: np.ndarray, device: torch.device, mode: str, input_size: int = 512, mask2_thr: float = 0.5):
+    h0, w0 = img_bgr.shape[:2]
+    x = preprocess_img(img_bgr, input_size, device)
+
+    out = base_model(pixel_values=x, output_hidden_states=True)
+    logits_seg = out.logits
+    pred_ids = torch.argmax(logits_seg, dim=1)[0].detach().cpu().numpy().astype(np.uint8)
+    pred_ids_full = cv2.resize(pred_ids, (w0, h0), interpolation=cv2.INTER_NEAREST)
+
+    result = {
+        "pred_seg": pred_ids_full,
+        "mask2_prob": None,
+        "mask2_bin": None,
+        "label_id": None,
+        "label_conf": None,
+        "label_probs": None,
+    }
+
+    if mode == "single" or aux_head is None:
+        return result
+
+    feat = out.hidden_states[-1]
+    if feat.shape[-2:] != logits_seg.shape[-2:]:
+        feat = F.interpolate(feat, size=logits_seg.shape[-2:], mode="bilinear", align_corners=False)
+
+    if mode == "mask2":
+        logits_m2 = aux_head(feat, logits_seg)
+        prob = torch.sigmoid(logits_m2)[0, 0].detach().float().cpu().numpy()
+        bin2 = (prob >= mask2_thr).astype(np.uint8) * 255
+        result["mask2_prob"] = cv2.resize(prob, (w0, h0), interpolation=cv2.INTER_LINEAR)
+        result["mask2_bin"] = cv2.resize(bin2, (w0, h0), interpolation=cv2.INTER_NEAREST)
+        return result
+
+    if mode == "label":
+        logits_label = aux_head(feat, logits_seg)
+        probs = torch.softmax(logits_label, dim=1)[0].detach().cpu().numpy()
+        lid = int(np.argmax(probs))
+        result["label_id"] = lid
+        result["label_conf"] = float(probs[lid])
+        result["label_probs"] = probs
+        return result
+
+    return result
+
+
+# ============================================================
+# Painéis
+# ============================================================
+
+def make_label_prob_bar(probs: np.ndarray, names: Dict[int, str], width: int = 520, row_h: int = 34) -> np.ndarray:
+    n = len(probs)
+    h = max(row_h * n, row_h)
+    img = np.zeros((h, width, 3), dtype=np.uint8) + 35
+    for i, p in enumerate(probs):
+        y0 = i * row_h
+        bar_w = int((width - 180) * float(p))
+        cv2.rectangle(img, (170, y0 + 8), (170 + bar_w, y0 + row_h - 8), (80, 180, 80), -1)
+        txt = f"{i} {names.get(i, f'label_{i}')}: {p:.3f}"
+        cv2.putText(img, txt, (8, y0 + 23), cv2.FONT_HERSHEY_SIMPLEX, 0.55, (245, 245, 245), 1, cv2.LINE_AA)
+    return img
+
+
+def make_panel_single(sample: Sample, lut_bgr, img_bgr, gt1, pred1, segm, id_to_name):
+    pred_col = colorize_seg(pred1, lut_bgr)
+    pred_ov = overlay_mask(img_bgr, pred_col, 0.45)
+    if gt1 is not None:
+        gt_col = colorize_seg(gt1, lut_bgr)
+        gt_ov = overlay_mask(img_bgr, gt_col, 0.45)
+    else:
+        gt_ov = img_bgr.copy()
+
+    tile_w = 520
+    row = np.concatenate([resize_fit(img_bgr, tile_w), resize_fit(gt_ov, tile_w), resize_fit(pred_ov, tile_w)], axis=1)
+    lines = [
+        f"{sample.group_name}/{sample.filename}",
+        "Modo: single | D/A navega | Q sai",
+    ]
+    if segm:
+        lines.append(f"SEG acc={segm['acc']:.3f} mIoU={segm['miou']:.3f}")
+    lines.append("IMG | GT SEG | PRED SEG")
+    return put_hud(row, lines)
+
+
+def make_panel_mask2(sample: Sample, lut_bgr, img_bgr, gt1, pred1, gt2, prob2, bin2, mask2_thr, segm, m2m):
+    pred_col = colorize_seg(pred1, lut_bgr)
+    pred_ov = overlay_mask(img_bgr, pred_col, 0.45)
+    if gt1 is not None:
+        gt_col = colorize_seg(gt1, lut_bgr)
+        gt_ov = overlay_mask(img_bgr, gt_col, 0.45)
+    else:
+        gt_ov = img_bgr.copy()
+
+    prob = np.clip(prob2, 0.0, 1.0).astype(np.float32)
+    g = (prob * 255.0).astype(np.uint8)
+    r = ((1.0 - prob) * 255.0).astype(np.uint8)
+    b = np.zeros_like(g, dtype=np.uint8)
+    prob_solid = cv2.merge([b, g, r])
+
+    edges = cv2.Canny(bin2, 50, 150)
+    prob_solid[edges > 0] = (255, 255, 255)
+    bin2_solid = cv2.cvtColor(bin2, cv2.COLOR_GRAY2BGR)
+
+    if gt2 is not None:
+        gt2_vis = gt2.copy().astype(np.uint8)
+        gt2_vis[gt2_vis > 1] = 255
+        gt2_bgr = cv2.cvtColor(gt2_vis, cv2.COLOR_GRAY2BGR)
+        gt2_ov = overlay_mask(img_bgr, gt2_bgr, 0.35)
+    else:
+        gt2_ov = img_bgr.copy()
+
+    tile_w = 520
+    tile_h = int(tile_w * img_bgr.shape[0] / img_bgr.shape[1])
+    row1 = np.concatenate([resize_fit(img_bgr, tile_w, tile_h), resize_fit(gt_ov, tile_w, tile_h), resize_fit(pred_ov, tile_w, tile_h)], axis=1)
+    row2 = np.concatenate([resize_fit(gt2_ov, tile_w, tile_h), resize_fit(prob_solid, tile_w, tile_h), resize_fit(bin2_solid, tile_w, tile_h)], axis=1)
+    panel = np.concatenate([row1, row2], axis=0)
+
+    lines = [
+        f"{sample.group_name}/{sample.filename}",
+        f"Modo: dual_head_mask | thr={mask2_thr:.2f} | W/S threshold | D/A navega | Q sai",
+    ]
+    if segm:
+        lines.append(f"SEG acc={segm['acc']:.3f} mIoU={segm['miou']:.3f}")
+    if m2m:
+        lines.append(f"MASK2 acc={m2m['acc']:.3f} IoU={m2m['iou']:.3f}")
+    lines.append("Topo: IMG | GT SEG | PRED SEG")
+    lines.append("Baixo: GT MASK2 | PRED MASK2 prob | PRED MASK2 bin")
+    return put_hud(panel, lines)
+
+
+def make_panel_label(sample: Sample, lut_bgr, img_bgr, gt1, pred1, pred_label_id, pred_conf, label_probs, gt_label_id, gt_label_name, label_names, segm):
+    pred_col = colorize_seg(pred1, lut_bgr)
+    pred_ov = overlay_mask(img_bgr, pred_col, 0.45)
+    if gt1 is not None:
+        gt_col = colorize_seg(gt1, lut_bgr)
+        gt_ov = overlay_mask(img_bgr, gt_col, 0.45)
+    else:
+        gt_ov = img_bgr.copy()
+
+    tile_w = 520
+    tile_h = int(tile_w * img_bgr.shape[0] / img_bgr.shape[1])
+    row1 = np.concatenate([resize_fit(img_bgr, tile_w, tile_h), resize_fit(gt_ov, tile_w, tile_h), resize_fit(pred_ov, tile_w, tile_h)], axis=1)
+
+    probs_panel = make_label_prob_bar(label_probs, label_names, width=tile_w)
+    blank = np.zeros_like(probs_panel) + 25
+    row2 = np.concatenate([probs_panel, blank, blank], axis=1)
+    panel = np.concatenate([row1, row2], axis=0)
+
+    pred_name = label_names.get(pred_label_id, f"label_{pred_label_id}") if pred_label_id is not None else "-"
+    gt_name = gt_label_name or (label_names.get(gt_label_id, f"label_{gt_label_id}") if gt_label_id is not None else "-")
+    ok_txt = "OK" if (gt_label_id is not None and pred_label_id == gt_label_id) else "MISS" if gt_label_id is not None else "SEM_GT"
+
+    lines = [
+        f"{sample.group_name}/{sample.filename}",
+        "Modo: dual_head_label | D/A navega | Q sai",
+        f"GT={gt_name} ({gt_label_id}) | PRED={pred_name} ({pred_label_id}) conf={pred_conf:.3f} | {ok_txt}",
+    ]
+    if segm:
+        lines.append(f"SEG acc={segm['acc']:.3f} mIoU={segm['miou']:.3f}")
+    lines.append("Topo: IMG | GT SEG | PRED SEG | Baixo: probabilidades por estado")
+    return put_hud(panel, lines)
+
+
+def compute_roi_indices(h: int, roi_inicio: float, roi_tamanho: float):
+    y0 = int(h * roi_inicio)
+    y1 = int(h * min(1.0, roi_inicio + roi_tamanho))
+    y0 = max(0, min(h - 1, y0))
+    y1 = max(y0 + 1, min(h, y1))
+    return y0, y1
+
+
+def resize_to_config(img_bgr: np.ndarray, resolucao):
+    w, h = int(resolucao[0]), int(resolucao[1])
+    return cv2.resize(img_bgr, (w, h), interpolation=cv2.INTER_AREA)
+
+
+def draw_label_result(img, label_id, label_conf, label_probs, label_names):
+    if label_id is None:
+        return img
+
+    name = label_names.get(label_id, f"label_{label_id}")
+    txt = f"STATUS: {name} | conf={label_conf:.3f}"
+    cv2.putText(img, txt, (12, 64), cv2.FONT_HERSHEY_SIMPLEX, 0.75, (20, 20, 20), 4, cv2.LINE_AA)
+    cv2.putText(img, txt, (12, 64), cv2.FONT_HERSHEY_SIMPLEX, 0.75, (245, 245, 245), 2, cv2.LINE_AA)
+    return img
+
+
+# ============================================================
+# Main
+# ============================================================
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--config", type=str, default="config.json")
+    ap.add_argument("--ckpt", type=str, default=None)
+    ap.add_argument("--split", default="test", choices=["test", "val", "train"])
+    ap.add_argument("--input_size", type=int, default=512)
+    ap.add_argument("--mask2_thr", type=float, default=0.5)
+    ap.add_argument("--save_dir", type=str, default=None, help="Se informado, salva painéis em disco além de abrir viewer.")
+    ap.add_argument("--no_view", action="store_true", help="Não abre janela, apenas salva se --save_dir for informado e imprime métricas.")
+    ap.add_argument("--norm_stats", type=str, default=None, help="Caminho para JSON com mean/std por canal (ex: norm_stats.json).")
+    ap.add_argument("--test_folder", type=str, default=None, help="pasta para teste")
+    ap.add_argument("--camera", action="store_true", help="Inferir em tempo real pela OAK-D Lite")
+    ap.add_argument("--camera_fps", type=int, default=20)
+    ap.add_argument("--camera_res", type=str, default="720p", choices=["720p", "1080p"])
+    args = ap.parse_args()
+
+    device = find_device()
+    print("Device:", device)
+
+    with open(args.config, "r", encoding="utf-8") as f:
+        config = json.load(f)
+
+    MODELO = config["camera"]
+    BACKBONE = config["backbone"]
+    RESOLUCAO = config.get("resolucao", [args.input_size, args.input_size])
+    ROI_INICIO = float(config.get("roi_inicio", 0.0))
+    ROI_TAMANHO = float(config.get("roi_tamanho", 1.0))
+    dataset_path = os.path.join(MODELO, "dataset")
+    mode = infer_mode_from_config(config)
+    print(f"[INFO] mode={mode}")
+
+    if args.camera:
+        samples = []
+        print("[INFO] modo camera OAK-D Lite")
+    elif args.test_folder:
+        samples = discover_image_folder(args.test_folder)
+        print(f"[INFO] samples: {len(samples)} | test_folder={args.test_folder}")
+    else:
+        split_root = os.path.join(dataset_path, "split", args.split)
+        if not os.path.isdir(split_root):
+            print(f"[WARN] split/{args.split} não existe. Vou usar split/val.")
+            split_root = os.path.join(dataset_path, "split", "val")
+
+        samples = discover_samples(split_root)
+        print(f"[INFO] samples: {len(samples)} | split_root={split_root}")
+
+    labelmap_path = os.path.join(dataset_path, "labelmap.txt")
+    lut_bgr, id_to_name, ignore_id = load_labelmap_lut_bgr(labelmap_path)
+    num_classes = len(id_to_name)
+    print("[labelmap]", id_to_name)
+
+    save_path = default_save_path(config, mode)
+    if args.ckpt:
+        ckpt_path = args.ckpt
+    else:
+        # Para label, costuma ser interessante ver best_label; para mask2, best_mask2; fallback best_miou
+        candidates = []
+        if mode == "label":
+            candidates += [os.path.join(save_path, "best_label.pt")]
+        elif mode == "mask2":
+            candidates += [os.path.join(save_path, "best_mask2.pt")]
+        candidates += [os.path.join(save_path, "best_miou.pt"), os.path.join(save_path, "last.pt")]
+        ckpt_path = next((p for p in candidates if os.path.isfile(p)), candidates[0])
+
+    if not os.path.isfile(ckpt_path):
+        raise SystemExit(f"Checkpoint não encontrado: {ckpt_path}")
+    print(f"[model] ckpt={ckpt_path}")
+
+    base_model = build_base_model(num_classes=num_classes, device=device, backbone=BACKBONE)
+
+    aux_head = None
+    label_names = {}
+    feat_ch = get_feat_ch(base_model, device, input_h=args.input_size, input_w=args.input_size)
+
+    # Carrega ckpt antes para descobrir metadados se necessário
+    ckpt_pre = torch.load(ckpt_path, map_location="cpu", weights_only=False)
+    extra = ckpt_pre.get("extra", {}) or {}
+
+    if mode == "mask2":
+        aux_head = Mask2Head(feat_ch=feat_ch, num_classes=num_classes, hidden=256, dropout=0.1)
+    elif mode == "label":
+        print(extra)
+        label_names_raw = extra.get("label_name_by_id", {}) or {}
+        # JSON pode salvar keys como string
+        label_names = {int(k): str(v) for k, v in label_names_raw.items()} if label_names_raw else {}
+        print(label_names_raw)
+
+        # Se não tiver no ckpt, tenta inferir do dataset de labels
+        max_label_id = -1
+        for s in samples:
+            lid, lname = read_gt_label(s)
+            if lid is not None:
+                max_label_id = max(max_label_id, int(lid))
+                if lname:
+                    label_names[int(lid)] = str(lname)
+        if max_label_id < 0:
+            # último fallback: pelo shape da última camada do checkpoint
+            sd = ckpt_pre.get("aux_head", {})
+            for k, v in sd.items():
+                if k.endswith("net.3.weight") or k.endswith("net.3.bias"):
+                    max_label_id = int(v.shape[0]) - 1
+                    break
+        num_label_classes = max_label_id + 1
+        if num_label_classes <= 0:
+            raise RuntimeError("Não consegui inferir num_label_classes para LabelHead.")
+        for i in range(num_label_classes):
+            label_names.setdefault(i, f"label_{i}")
+        aux_head = LabelHead(feat_ch=feat_ch, num_seg_classes=num_classes, num_label_classes=num_label_classes, hidden=256, dropout=0.2)
+        print(f"[label_head] classes={num_label_classes} names={label_names}")
+
+    ckpt = load_checkpoint(ckpt_path, base_model, aux_head, device)
+    print(f"[ckpt] epoch={ckpt.get('epoch')} bests={ckpt.get('bests')}")
+
+    if args.save_dir:
+        os.makedirs(args.save_dir, exist_ok=True)
+
+    # ==========================
+    # Normalizador (fixo ou dinâmico)
+    # ==========================
+    global NORM_MEAN, NORM_STD
+    norm_stats = None
+
+    # Caminho padrão: dentro do dataset, nome do arquivo de stats
+    # (ajusta aqui pro nome que você realmente usou: norm_stats.json, por ex.)
+    norm_stats_path = os.path.join(save_path, "norm_stats.json")
+    if args.norm_stats is not None:
+        norm_stats_path = args.norm_stats
+
+    if norm_stats_path is not None and os.path.exists(norm_stats_path):
+        with open(norm_stats_path, "r", encoding="utf-8") as f:
+            norm_stats = json.load(f)
+
+        stats_channels = norm_stats.get("channels", [])
+        stats_mean = norm_stats.get("mean", [])
+        stats_std  = norm_stats.get("std", [])
+
+        print(f"[NORM] usando stats fixos de: {norm_stats_path}")
+        print(f"[NORM] channels={stats_channels}")
+        print(f"[NORM] mean={stats_mean}")
+        print(f"[NORM] std ={stats_std}")
+
+        # Garante que temos pelo menos R,G,B
+        idx_by_name = {name: i for i, name in enumerate(stats_channels)}
+        required = ["R", "G", "B"]
+        if not all(ch in idx_by_name for ch in required):
+            print("[NORM] AVISO: norm_stats não contém todos os canais R,G,B. Mantendo normalize imagenet.")
+        else:
+            NORM_MEAN = torch.tensor(stats_mean, dtype=torch.float32, device=device).view(3, 1, 1)
+            NORM_STD = torch.tensor(stats_std, dtype=torch.float32, device=device).view(3, 1, 1).clamp_min(1e-6)
+
+            print("[NORM] Normalização fixa por canal ativada para [R,G,B].")
+    else:
+        if norm_stats_path:
+            print(f"[NORM] Caminho de norm_stats não encontrado: {norm_stats_path}. Usando normalize imagenet.")
+        else:
+            print("[NORM] norm_stats não informado. Usando normalize imagenet.")
+
+    if args.camera:
+        pipeline = dai.Pipeline()
+
+        cam_rgb = pipeline.create(dai.node.Camera).build()
+
+        rgb_out = cam_rgb.requestOutput(
+            size=(RESOLUCAO[0], RESOLUCAO[1]),
+            type=dai.ImgFrame.Type.BGR888p,
+            fps=args.camera_fps
+        )
+
+        rgb_queue = rgb_out.createOutputQueue(maxSize=4, blocking=False)
+
+        win = "SegFormer Dual - OAK-D Lite"
+        cv2.namedWindow(win, cv2.WINDOW_NORMAL)
+
+        pipeline.start()
+
+        prev_time = time.time()
+
+        while pipeline.isRunning():
+            in_rgb = rgb_queue.get()
+            frame_bgr = in_rgb.getCvFrame()
+
+            H, W = frame_bgr.shape[:2]
+            y0, y1 = compute_roi_indices(H, ROI_INICIO, ROI_TAMANHO)
+
+            roi_bgr = frame_bgr[y0:y1, 0:W]
+            roi_input = resize_to_config(roi_bgr, RESOLUCAO)
+
+            res = infer_sample(
+                base_model=base_model,
+                aux_head=aux_head,
+                img_bgr=roi_input,
+                device=device,
+                mode=mode,
+                input_size=args.input_size,
+                mask2_thr=args.mask2_thr,
+            )
+
+            pred_seg = res["pred_seg"]
+            pred_col = colorize_seg(pred_seg, lut_bgr)
+            pred_col_roi = cv2.resize(
+                pred_col,
+                (roi_bgr.shape[1], roi_bgr.shape[0]),
+                interpolation=cv2.INTER_NEAREST
+            )
+
+            overlay = frame_bgr.copy()
+            overlay[y0:y1, 0:W] = overlay_mask(
+                overlay[y0:y1, 0:W],
+                pred_col_roi,
+                alpha=0.45
+            )
+
+            now = time.time()
+            fps = 1.0 / max(1e-6, now - prev_time)
+            prev_time = now
+
+            lines = [
+                f"FPS: {fps:.1f}",
+                f"Modo: {mode}",
+                "Q/ESC sair",
+            ]
+
+            if mode == "label":
+                lid = res["label_id"]
+                conf = res["label_conf"] or 0.0
+                lname = label_names.get(lid, f"label_{lid}") if lid is not None else "-"
+                lines.append(f"Label: {lname} ({lid}) conf={conf:.3f}")
+
+            elif mode == "mask2" and res["mask2_bin"] is not None:
+                mask2_bin = cv2.resize(
+                    res["mask2_bin"],
+                    (roi_bgr.shape[1], roi_bgr.shape[0]),
+                    interpolation=cv2.INTER_NEAREST
+                )
+                edges = cv2.Canny(mask2_bin, 50, 150)
+                overlay[y0:y1, 0:W][edges > 0] = (255, 255, 255)
+                lines.append(f"Mask2 thr={args.mask2_thr:.2f}")
+
+            overlay = put_hud(overlay, lines)
+
+            cv2.imshow(win, overlay)
+            k = cv2.waitKey(1) & 0xFF
+
+            if k in (ord("q"), ord("Q"), 27):
+                break
+
+            if mode == "mask2" and k in (ord("w"), ord("W")):
+                args.mask2_thr = min(0.95, args.mask2_thr + 0.05)
+
+            if mode == "mask2" and k in (ord("s"), ord("S")):
+                args.mask2_thr = max(0.05, args.mask2_thr - 0.05)
+
+        cv2.destroyAllWindows()
+        return
+
+    idx = 0
+    win = "SegFormer Test Viewer"
+    if not args.no_view:
+        cv2.namedWindow(win, cv2.WINDOW_NORMAL)
+
+    global_seg_ious = []
+    global_seg_accs = []
+    global_mask2_iou = []
+    global_mask2_acc = []
+    label_total = 0
+    label_correct = 0
+
+    while True:
+        s = samples[idx]
+        img = imread_bgr(s.img_path)
+        gt1 = imread_gray(s.mask_path) if s.mask_path else None
+        gt2 = imread_gray(s.mask2_path) if s.mask2_path else None
+        gt_label_id, gt_label_name = read_gt_label(s)
+
+        res = infer_sample(
+            base_model=base_model,
+            aux_head=aux_head,
+            img_bgr=img,
+            device=device,
+            mode=mode,
+            input_size=args.input_size,
+            mask2_thr=args.mask2_thr,
+        )
+
+        pred1 = res["pred_seg"]
+        segm = seg_metrics(pred1, gt1, num_classes=num_classes, ignore_id=ignore_id)
+        if segm:
+            global_seg_ious.append(segm["miou"])
+            global_seg_accs.append(segm["acc"])
+
+        if mode == "mask2":
+            m2m = mask2_metrics(res["mask2_bin"], gt2)
+            if m2m:
+                global_mask2_iou.append(m2m["iou"])
+                global_mask2_acc.append(m2m["acc"])
+            panel = make_panel_mask2(
+                sample=s,
+                lut_bgr=lut_bgr,
+                img_bgr=img,
+                gt1=gt1,
+                pred1=pred1,
+                gt2=gt2,
+                prob2=res["mask2_prob"],
+                bin2=res["mask2_bin"],
+                mask2_thr=args.mask2_thr,
+                segm=segm,
+                m2m=m2m,
+            )
+        elif mode == "label":
+            pred_label_id = res["label_id"]
+            pred_conf = res["label_conf"] if res["label_conf"] is not None else 0.0
+            if gt_label_id is not None:
+                label_total += 1
+                if pred_label_id == gt_label_id:
+                    label_correct += 1
+            panel = make_panel_label(
+                sample=s,
+                lut_bgr=lut_bgr,
+                img_bgr=img,
+                gt1=gt1,
+                pred1=pred1,
+                pred_label_id=pred_label_id,
+                pred_conf=pred_conf,
+                label_probs=res["label_probs"],
+                gt_label_id=gt_label_id,
+                gt_label_name=gt_label_name,
+                label_names=label_names,
+                segm=segm,
+            )
+        else:
+            panel = make_panel_single(s, lut_bgr, img, gt1, pred1, segm, id_to_name)
+
+        if args.save_dir:
+            out_name = f"{idx:05d}_{s.group_name.replace('/', '_')}_{os.path.splitext(s.filename)[0]}.jpg"
+            cv2.imwrite(os.path.join(args.save_dir, out_name), panel)
+
+        if args.no_view:
+            idx += 1
+            if idx >= len(samples):
+                break
+            continue
+
+        cv2.imshow(win, panel)
+        k = cv2.waitKey(0) & 0xFF
+
+        if k in (ord("q"), ord("Q"), 27):
+            break
+        if k in (ord("a"), ord("A"), 81):
+            idx = (idx - 1) % len(samples)
+            continue
+        if k in (ord("d"), ord("D"), 83):
+            idx = (idx + 1) % len(samples)
+            continue
+        if mode == "mask2" and k in (ord("w"), ord("W")):
+            args.mask2_thr = min(0.95, args.mask2_thr + 0.05)
+            continue
+        if mode == "mask2" and k in (ord("s"), ord("S")):
+            args.mask2_thr = max(0.05, args.mask2_thr - 0.05)
+            continue
+
+    if not args.no_view:
+        cv2.destroyAllWindows()
+
+    print("\nResumo parcial/geral da sessão:")
+    if global_seg_ious:
+        print(f"SEG  : mean_acc={np.mean(global_seg_accs):.4f} mean_mIoU={np.mean(global_seg_ious):.4f} n={len(global_seg_ious)}")
+    if mode == "mask2" and global_mask2_iou:
+        print(f"MASK2: mean_acc={np.mean(global_mask2_acc):.4f} mean_IoU={np.mean(global_mask2_iou):.4f} n={len(global_mask2_iou)}")
+    if mode == "label" and label_total > 0:
+        print(f"LABEL: acc={label_correct / max(1, label_total):.4f} correct={label_correct}/{label_total}")
+
+
+if __name__ == "__main__":
+    main()

Python/OAK/datasets/config.json

@@ -2,7 +2,8 @@
     "camera": "oak-d",
     "modelo": "segformer_b0",
     "model_name": "nav_mit",
-    "dual_head": false,
+    "dual_head_label": true,
+    "dual_head_mask": false,
     "main_class_name": "navegavel", 
     "es_classes": "", 
     "model_to_use": "geral",
@@ -13,5 +14,15 @@
     "shaves": 3,
     "channels": 3,
     "use_ndvi": false,
-    "backbone": "nvidia/mit-b0"
+    "backbone": "nvidia/mit-b0",
+    "label_classes": [
+        "Parado",
+        "EntrandoRua",
+        "CaminhandoRua",
+        "SaindoRua",
+        "Manobrando",
+        "Direcionando",
+        "RetornandoBase",
+        "Indefinido"
+    ]
 }

Python/OAK/datasets/download_backbone.py

@@ -0,0 +1,12 @@
+from transformers import SegformerForSemanticSegmentation, SegformerImageProcessor
+
+modelo = "nvidia/mit-b0"
+pasta = r"C:\AgroBaseModels\Backbones\nvidia_mit_b0"
+
+model = SegformerForSemanticSegmentation.from_pretrained(modelo)
+processor = SegformerImageProcessor.from_pretrained(modelo)
+
+model.save_pretrained(pasta)
+processor.save_pretrained(pasta)
+
+print("Backbone salvo em:", pasta)

Python/OAK/datasets/multiespec_module/multispectral_client.py

@@ -216,7 +216,23 @@ class MultiSpectralClient:
 
     def get_next_decoded(self, timeout=2.0, update_radiometry=True):
         frame, meta = self.get_next_frame(timeout=timeout)
-        decoded = self.core.decode_stream_cameras(frame, meta)
+
+        frame_type = meta.get("frame_type")
+
+        if frame_type == "RAW_BRUTO":
+            decoded = self.core.decode_stream_cameras(frame, meta)
+
+        elif frame_type in ("RGB", "MULTISPEC"):
+            decoded = {
+                "cam2": {
+                    "name": "RGB",
+                    "image": frame,
+                    "meta": meta
+                }
+            }
+
+        else:
+            raise RuntimeError(f"frame_type não suportado: {frame_type}")
 
         if update_radiometry:
             self.update_radiometry(decoded, meta)

Python/OAK/datasets/oak-fcc-3/_0_capture.py

@@ -230,6 +230,7 @@ def main():
 
     try:
         with MultiSpectralClient(
+            #mx_id="194430108133AC2F00",
             width=raw_w,
             height=raw_h,
             bayer=args.bayer,

Python/OAK/datasets/oak-fcc-3/_6_normalize.py

@@ -82,6 +82,43 @@ DEFAULT_DATASET_BASE = "dataset"
 CHANNEL_NAMES = ["R", "G", "B", "RE", "NIR"]
 
 
+_CORE_CACHE: Dict[tuple, RawProcessorCore] = {}
+
+def get_or_create_core(
+    raw_size: Tuple[int, int],
+    bayer_pattern: str,
+    module_params_path: Optional[str],
+) -> RawProcessorCore:
+    raw_w, raw_h = raw_size
+    bayer = str(bayer_pattern or "BGGR").upper()
+
+    key = (
+        int(raw_w),
+        int(raw_h),
+        bayer,
+        str(Path(module_params_path).resolve()) if module_params_path else None,
+    )
+
+    core = _CORE_CACHE.get(key)
+    if core is not None:
+        return core
+
+    print(
+        "[CORE_CACHE] criando RawProcessorCore "
+        f"raw={raw_w}x{raw_h} bayer={bayer} module={module_params_path}"
+    )
+
+    core = RawProcessorCore(
+        sensor_width=int(raw_w),
+        sensor_height=int(raw_h),
+        bayer_pattern=bayer,
+        calibration_json_path=module_params_path,
+    )
+
+    _CORE_CACHE[key] = core
+    return core
+
+
 @dataclass
 class SampleBundle:
     group: str
@@ -198,6 +235,13 @@ def collect_samples_from_group(group_dir: Path) -> List[SampleBundle]:
     return samples
 
 
+def copy_json_safe(obj):
+    try:
+        return json.loads(json.dumps(obj, default=str))
+    except Exception:
+        return None
+
+
 # ============================================================
 # Module params / tensor
 # ============================================================
@@ -304,14 +348,12 @@ def build_tensor_from_sample(
     raw_size: Tuple[int, int],
     module_params_path: Optional[str],
 ) -> Tuple[np.ndarray, dict]:
-    raw_w, raw_h = raw_size
-    bayer = str(meta.get("bayer_pattern") or meta.get("bayer") or "RGGB").upper()
+    bayer = str(meta.get("bayer_pattern") or meta.get("bayer") or "BGGR").upper()
 
-    core = RawProcessorCore(
-        sensor_width=int(raw_w),
-        sensor_height=int(raw_h),
+    core = get_or_create_core(
+        raw_size=raw_size,
         bayer_pattern=bayer,
-        calibration_json_path=module_params_path,
+        module_params_path=module_params_path,
     )
 
     frame = load_frame_from_saved_bins(sample, meta)
@@ -329,14 +371,15 @@ def build_tensor_from_sample(
     info = {
         "module_params": module_params_path,
         "bayer_pattern": bayer,
-        "fusion_result": getattr(core, "last_fusion_result", None),
-        "patch_normalization_result": getattr(core, "last_patch_normalization_result", None),
-        "frame_quality": getattr(core, "last_frame_quality_result", None),
+        "fusion_result": copy_json_safe(getattr(core, "last_fusion_result", None)),
+        "patch_normalization_result": copy_json_safe(getattr(core, "last_patch_normalization_result", None)),
+        "frame_quality": copy_json_safe(getattr(core, "last_frame_quality_result", None)),
     }
 
     return tensor, info
 
 
+
 # ============================================================
 # Máscara / preview
 # ============================================================
@@ -407,6 +450,82 @@ def save_bgr(path: Path, bgr: np.ndarray):
         raise RuntimeError(f"Falha ao salvar preview: {path}")
 
 
+def get_class_ids_from_config() -> dict:
+    heads = config.get("heads", {}) or {}
+    semantic = heads.get("semantic", {}) or {}
+    classes = semantic.get("classes", {}) or {}
+
+    return {
+        "chao": int(classes.get("chao", 0)),
+        "cana": int(classes.get("cana", 1)),
+        "erva": int(classes.get("erva", 2)),
+    }
+
+
+def build_head_masks_from_semantic(
+    mask_ids: np.ndarray,
+    class_ids: dict,
+    ignore_id: int = 255,
+) -> dict:
+    """
+    Gera máscaras auxiliares para multi-head a partir da máscara semântica alinhada.
+
+    Entrada:
+      mask_ids: HW com IDs semânticos.
+
+    Saída:
+      {
+        "semantic": HW uint8,
+        "vegetation": HW uint8 com 0/1/ignore,
+        "cana": HW uint8 com 0/1/ignore
+      }
+    """
+    chao_id = int(class_ids.get("chao", 0))
+    cana_id = int(class_ids.get("cana", 1))
+    erva_id = int(class_ids.get("erva", 2))
+
+    semantic = mask_ids.astype(np.uint8, copy=True)
+
+    ignore_mask = semantic == ignore_id
+
+    vegetation = np.zeros_like(semantic, dtype=np.uint8)
+    vegetation[(semantic == cana_id) | (semantic == erva_id)] = 1
+    vegetation[ignore_mask] = ignore_id
+
+    cana = np.zeros_like(semantic, dtype=np.uint8)
+    cana[semantic == cana_id] = 1
+    cana[ignore_mask] = ignore_id
+
+    return {
+        "semantic": semantic,
+        "vegetation": vegetation,
+        "cana": cana,
+    }
+
+
+def save_mask_npy_and_debug(mask: np.ndarray, npy_path: Path, png_path: Path):
+    ensure_dir(npy_path.parent)
+    ensure_dir(png_path.parent)
+
+    mask = mask.astype(np.uint8, copy=False)
+    np.save(str(npy_path), mask)
+
+    # Debug visual:
+    # 0 -> preto
+    # 1 -> cinza claro
+    # 2 -> mais claro, quando existir na semântica
+    # 255 -> branco
+    debug = mask.copy()
+    debug_vis = np.zeros_like(debug, dtype=np.uint8)
+
+    debug_vis[debug == 0] = 0
+    debug_vis[debug == 1] = 120
+    debug_vis[debug == 2] = 220
+    debug_vis[debug == 255] = 255
+
+    cv2.imwrite(str(png_path), debug_vis)
+
+
 # ============================================================
 # Stats
 # ============================================================
@@ -467,11 +586,22 @@ def process_group(
 
     out_group = output_root / group_name
     out_tensors = out_group / "tensors"
+
     out_masks = out_group / "masks"
+    out_masks_vegetation = out_group / "masks_vegetation"
+    out_masks_cana = out_group / "masks_cana"
+
     out_metas = out_group / "metas"
     out_previews = out_group / "previews"
 
-    for d in (out_tensors, out_masks, out_metas, out_previews):
+    for d in (
+        out_tensors,
+        out_masks,
+        out_masks_vegetation,
+        out_masks_cana,
+        out_metas,
+        out_previews,
+    ):
         ensure_dir(d)
 
     rows = []
@@ -515,16 +645,43 @@ def process_group(
 
             # Salva tensor e mask
             tensor_path = out_tensors / f"{sample.base}.npy"
-            mask_path = out_masks / f"{sample.base}.npy"
-            mask_debug_path = out_masks / f"{sample.base}.png"
             preview_path = out_previews / f"{sample.base}.png"
             meta_out_path = out_metas / f"{sample.base}.json"
-
             np.save(str(tensor_path), tensor)
-            np.save(str(mask_path), mask_ids)
 
-            # Debug visual da máscara em IDs, só para inspeção rápida.
-            cv2.imwrite(str(mask_debug_path), mask_ids)
+            class_ids = get_class_ids_from_config()
+            head_masks = build_head_masks_from_semantic(
+                mask_ids=mask_ids,
+                class_ids=class_ids,
+                ignore_id=ignore_id,
+            )
+
+            mask_path = out_masks / f"{sample.base}.npy"
+            mask_debug_path = out_masks / f"{sample.base}.png"
+
+            mask_vegetation_path = out_masks_vegetation / f"{sample.base}.npy"
+            mask_vegetation_debug_path = out_masks_vegetation / f"{sample.base}.png"
+
+            mask_cana_path = out_masks_cana / f"{sample.base}.npy"
+            mask_cana_debug_path = out_masks_cana / f"{sample.base}.png"
+
+            save_mask_npy_and_debug(
+                head_masks["semantic"],
+                mask_path,
+                mask_debug_path,
+            )
+
+            save_mask_npy_and_debug(
+                head_masks["vegetation"],
+                mask_vegetation_path,
+                mask_vegetation_debug_path,
+            )
+
+            save_mask_npy_and_debug(
+                head_masks["cana"],
+                mask_cana_path,
+                mask_cana_debug_path,
+            )
 
             preview_bgr = tensor_to_preview_bgr(tensor)
             save_bgr(preview_path, preview_bgr)
@@ -542,6 +699,38 @@ def process_group(
                 "saved_payload_type": "tensor_npy",
                 "saved_tensor_path": safe_rel(tensor_path, dataset_root),
                 "saved_mask_path": safe_rel(mask_path, dataset_root),
+                "head_masks": {
+                    "semantic": {
+                        "path": safe_rel(mask_path, dataset_root),
+                        "shape": list(head_masks["semantic"].shape),
+                        "classes": {
+                            "chao": int(class_ids["chao"]),
+                            "cana": int(class_ids["cana"]),
+                            "erva": int(class_ids["erva"]),
+                        },
+                        "ignore_index": int(ignore_id),
+                    },
+                    "vegetation": {
+                        "path": safe_rel(mask_vegetation_path, dataset_root),
+                        "shape": list(head_masks["vegetation"].shape),
+                        "classes": {
+                            "background": 0,
+                            "vegetation": 1,
+                        },
+                        "positive_from_semantic": ["cana", "erva"],
+                        "ignore_index": int(ignore_id),
+                    },
+                    "cana": {
+                        "path": safe_rel(mask_cana_path, dataset_root),
+                        "shape": list(head_masks["cana"].shape),
+                        "classes": {
+                            "not_cana": 0,
+                            "cana": 1,
+                        },
+                        "positive_from_semantic": ["cana"],
+                        "ignore_index": int(ignore_id),
+                    },
+                },
                 "saved_preview_path": safe_rel(preview_path, dataset_root),
                 "saved_payload_dtype": str(tensor.dtype),
                 "saved_payload_shape": list(tensor.shape),
@@ -575,6 +764,8 @@ def process_group(
                 "base": sample.base,
                 "tensor": str(tensor_path),
                 "mask": str(mask_path),
+                "mask_vegetation": str(mask_vegetation_path),
+                "mask_cana": str(mask_cana_path),
                 "meta": str(meta_out_path),
                 "preview": str(preview_path),
                 "quality_status": frame_quality.get("status") if isinstance(frame_quality, dict) else None,
@@ -593,7 +784,17 @@ def process_group(
 def write_manifest(path: Path, rows: List[dict]):
     ensure_dir(path.parent)
 
-    fieldnames = ["group", "base", "tensor", "mask", "meta", "preview", "quality_status"]
+    fieldnames = [
+        "group",
+        "base",
+        "tensor",
+        "mask",
+        "mask_vegetation",
+        "mask_cana",
+        "meta",
+        "preview",
+        "quality_status",
+    ]
 
     with path.open("w", newline="", encoding="utf-8") as f:
         w = csv.DictWriter(f, fieldnames=fieldnames)

Python/OAK/datasets/oak-fcc-3/_7_split.py

@@ -18,10 +18,15 @@ RESOLUCAO = tuple(config.get("resolucao"))
 
 TENSOR_EXT = ".npy"
 MASK_NPY_SUFFIX = ".npy"
+AUX_MASK_DIRS = [
+    "masks_vegetation",
+    "masks_cana",
+]
 
 RE_ORIGINAL_PREFIX = re.compile(r"^original_(.+)$", re.IGNORECASE)
 RE_AUGMENTED_FAMILY = re.compile(r"^augmented_(.+?)(?:_aug[a-zA-Z0-9]*_\d+)?$", re.IGNORECASE)
 RE_AUG_SUFFIX = re.compile(r"_aug[a-zA-Z0-9]*_\d+$", re.IGNORECASE)
+MULTI_HEAD = bool(config.get("multi_head", False))
 
 
 def garantir(p):
@@ -237,6 +242,46 @@ def copiar_optional(src_dir, dst_dir, base, ext):
     return dst
 
 
+def copiar_mask_dir_optional(src_group_dir, dst_group_dir, mask_dir_name, base):
+    """
+    Copia uma pasta auxiliar de máscara, como:
+      masks_vegetation/
+      masks_cana/
+
+    Copia:
+      <base>.npy obrigatório se existir
+      <base>.png opcional se existir
+
+    Retorna caminhos de destino ou None.
+    """
+    src_dir = os.path.join(src_group_dir, mask_dir_name)
+    dst_dir = os.path.join(dst_group_dir, mask_dir_name)
+
+    npy_src = os.path.join(src_dir, base + ".npy")
+    png_src = os.path.join(src_dir, base + ".png")
+
+    result = {
+        "npy": None,
+        "png": None,
+    }
+
+    if not os.path.exists(npy_src):
+        return result
+
+    garantir(dst_dir)
+
+    npy_dst = os.path.join(dst_dir, base + ".npy")
+    shutil.copy2(npy_src, npy_dst)
+    result["npy"] = npy_dst
+
+    if os.path.exists(png_src):
+        png_dst = os.path.join(dst_dir, base + ".png")
+        shutil.copy2(png_src, png_dst)
+        result["png"] = png_dst
+
+    return result
+
+
 def copiar(
     nomes,
     src_group_dir,
@@ -304,11 +349,37 @@ def copiar(
                     shutil.copy2(cand, preview_dst)
                     break
 
+        aux_masks = {}
+        for aux_dir in AUX_MASK_DIRS:
+            aux_masks[aux_dir] = copiar_mask_dir_optional(
+                src_group_dir=src_group_dir,
+                dst_group_dir=dst_group_dir,
+                mask_dir_name=aux_dir,
+                base=base,
+            )
+
+        if MULTI_HEAD:
+            for aux_dir in AUX_MASK_DIRS:
+                aux_npy = aux_masks.get(aux_dir, {}).get("npy")
+                if aux_npy is None:
+                    raise RuntimeError(
+                        f"multi_head=true, mas máscara auxiliar ausente: "
+                        f"{aux_dir}/{base}.npy em {src_group_dir}"
+                    )
+
         rows.append({
             "base": base,
             "tensor": tensor_dst,
+
             "mask_npy": mask_npy_dst,
             "mask_png": mask_png_dst,
+
+            "mask_vegetation_npy": aux_masks.get("masks_vegetation", {}).get("npy"),
+            "mask_vegetation_png": aux_masks.get("masks_vegetation", {}).get("png"),
+
+            "mask_cana_npy": aux_masks.get("masks_cana", {}).get("npy"),
+            "mask_cana_png": aux_masks.get("masks_cana", {}).get("png"),
+
             "meta": meta_dst,
             "preview": preview_dst,
         })
@@ -447,8 +518,16 @@ def write_manifest(path, rows):
         "group",
         "base",
         "tensor",
+
         "mask_npy",
         "mask_png",
+
+        "mask_vegetation_npy",
+        "mask_vegetation_png",
+
+        "mask_cana_npy",
+        "mask_cana_png",
+
         "meta",
         "preview",
     ]

Python/OAK/datasets/oak-fcc-3/calibration/module_params.json

@@ -9,7 +9,8 @@
   "sensor_height": 800,
   "bayer_pattern": "BGGR",
   "rgb_processing": {
-    "mode": "bayer_planes"
+    "mode": "linear_demosaic",
+    "demosaic_algorithm": "bilinear"
   },
   "camera_settings": {
     "rgb": {
@@ -38,6 +39,9 @@
     }
   },
   "fusion_config": {
+    "use_remap_cache": true,
+    "use_remap_for_rgb": false,
+    "use_remap_for_spec": true,
     "alignment_mode": "homography",
     "baseline_mm": 75.0,
     "reference_camera": "rgb",
@@ -49,10 +53,11 @@
       },
       "nir": {
         "dx": 0,
-        "dy": 1,
+        "dy": 0,
         "theta_deg": 0.0
       }
     },
+    "homography_calibration_size": [1280, 800],
     "homographies": {
       "re_to_rgb": [
         [
@@ -91,7 +96,7 @@
     },
     "crop_valid_common": true,
     "resize_after_crop": true,
-    "target_size": null
+    "target_size": [1024,640]
   },
   "radiometric_config": {
     "enabled": false,
@@ -583,35 +588,35 @@
     "reference_mode": "fixed",
     "reference_controls": {
       "rgb": {
-        "exposure_time_us": 10000,
-        "sensitivity_iso": 400
+        "exposure_time_us": 2200,
+        "sensitivity_iso": 100
       },
       "re": {
-        "exposure_time_us": 15000,
-        "sensitivity_iso": 400
+        "exposure_time_us": 2500,
+        "sensitivity_iso": 100
       },
       "nir": {
-        "exposure_time_us": 15000,
-        "sensitivity_iso": 400
+        "exposure_time_us": 2500,
+        "sensitivity_iso": 100
       }
     },
 
     "scale_limits": {
       "default": {
         "min": 0.15,
-        "max": 6.0
+        "max": 3.0
       },
       "rgb": {
         "min": 0.15,
-        "max": 6.0
+        "max": 3.0
       },
       "re": {
         "min": 0.15,
-        "max": 8.0
+        "max": 2.5
       },
       "nir": {
         "min": 0.15,
-        "max": 8.0
+        "max": 2.5
       }
     },
 
@@ -676,21 +681,21 @@
   "rgb_calibration": {
     "enabled": true,
     "gains": {
-      "R": 1.2500000000000002,
+      "R": 1.061,
       "G": 1.0,
-      "B": 1.5500000000000005
+      "B": 1.452
     }
   },
   "flatfield_config": {
     "enabled": true,
-    "subtract_dark": true,
+    "subtract_dark": false,
     "schema": "multispec_flatfield_v1",
     "created_at": "2026-05-08 15:26:18",
     "json_file": "calibration/flatfield_maps_v1.json",
     "npz_file": "calibration/flatfield_maps_v1.npz",
-    "apply_before_fusion": true,
-    "apply_after_decode": true,
-    "apply_space": "native_camera_space",
+    "apply_before_fusion": false,
+    "apply_after_decode": false,
+    "apply_space": "final_tensor_space",
     "map_type": "gain",
     "formula": "channel_corrected = max(channel_linear - dark, 0) * gain_map",
     "channels": [
@@ -772,28 +777,30 @@
         "gain_std": 0.23178784549236298
       }
     },
-    "exp_gain_correct_during_flat_capture": false,
+    "exp_gain_correct_during_flat_capture": true,
     "smooth_ksize": 31,
     "min_gain": 0.25,
     "max_gain": 4.0,
     "notes": "",
-    "strength": 0.35,
+    "strength": 0.25,
     "strength_by_channel": {
-      "R": 0.9,
-      "G": 0.9,
-      "B": 0.9,
+      "R": 0.5,
+      "G": 0.5,
+      "B": 0.5,
       "RE": 0.25,
       "NIR": 0.25
     },
-
     "gain_min_runtime": 0.75,
     "gain_max_runtime": 1.35,
     "runtime_smooth_ksize": 81,
 
-    "saturation_guard_enabled": true,
+    "fast_runtime": true,
+    "saturation_guard_enabled": false,
     "saturation_guard_mode": "fade_strength",
     "saturation_guard_threshold": 0.97,
     "saturation_guard_soft_start": 0.88,
-    "saturation_guard_hard": 0.97
+    "saturation_guard_hard": 0.97,
+
+    "clip_output": true
   }
 }

Python/OAK/datasets/oak-fcc-3/config.json

@@ -1,8 +1,7 @@
 {
     "camera": "oak-fcc-3",
     "modelo": "segformer_b1",
-    "model_name": "test_1",
-    "dual_head": false,
+    "model_name": "target_teached",
     "main_class_name": "cana", 
     "es_classes": "", 
     "model_to_use": "geral",
@@ -12,9 +11,69 @@
     "roi_tamanho": 1.0,
     "shaves": 3,
     "channels": 5,
+    "input_channels": ["R", "G", "B", "RE", "NIR"],
     "use_ndvi": false,
     "backbone": "nvidia/mit-b1",
     "fusion_mode": "stacked",
-    "stats_source_tag": "stacked_raw5",
-    "module_params_json": "calibration/module_params.json"
+    "stats_source_tag": "stacked_raw4",
+    "module_params_json": "calibration/module_params.json",
+    "multi_head": true,
+    "heads": {
+        "semantic": {
+            "enabled": true,
+            "type": "multiclass",
+            "num_classes": 3,
+            "mask_dir": "masks",
+            "classes": {"chao": 0, "cana": 1, "erva": 2},
+            "ignore_index": 255,
+            "loss_weight": 0.10
+        },
+        "vegetation": {
+            "enabled": true,
+            "type": "binary",
+            "num_classes": 2,
+            "mask_dir": "masks_vegetation",
+            "classes": {"background": 0, "vegetation": 1},
+            "ignore_index": 255,
+            "loss_weight": 0.20
+        },
+        "cana": {
+            "enabled": true,
+            "type": "binary",
+            "num_classes": 2,
+            "mask_dir": "masks_cana",
+            "classes": {"not_cana": 0, "cana": 1},
+            "ignore_index": 255,
+            "loss_weight": 0.25
+        },
+        "target": {
+            "enabled": true,
+            "type": "binary",
+            "num_classes": 2,
+            "mask_dir": "__derived_target__",
+            "classes": {"background": 0, "target": 1},
+            "ignore_index": 255,
+            "loss_weight": 0.45,
+            "derived": true
+        }
+    },
+    "target_distillation": {
+        "enabled": true,
+
+        "hard_weight": 0.85,
+        "distill_weight": 0.15,
+
+        "rampup_enabled": true,
+        "start_epoch": 8,
+        "rampup_epochs": 12,
+
+        "w_sem_erva": 0.45,
+        "w_veg_not_cana": 0.35,
+        "w_veg_suppressed": 0.20,
+        "cana_suppression_power": 1.5,
+
+        "teacher_min": 0.0,
+        "teacher_max": 1.0,
+        "detach_teacher": true
+    }
 }

Python/OAK/datasets/oak-fcc-3/core/benchmark.py

@@ -0,0 +1,620 @@
+import argparse
+import json
+import sys
+import time
+from pathlib import Path
+
+import cv2
+import numpy as np
+
+try:
+    import torch
+except Exception:
+    torch = None
+
+
+# ============================================================
+# Ajuste de import local
+# ============================================================
+
+THIS_FILE = Path(__file__).resolve()
+
+# Esperado:
+# .../Python/Scripts/workers/camera_worker/oak_fcc3_core/benchmark_raw_bruto_scientific.py
+WORKERS_DIR = THIS_FILE.parents[2]
+
+if str(WORKERS_DIR) not in sys.path:
+    sys.path.insert(0, str(WORKERS_DIR))
+
+from camera_worker.oak_fcc3_core.oak_fcc3_client import OakFcc3Client
+
+try:
+    from camera_worker.oak_fcc3_core.segformer_service import MultiSpecSegformerService
+except Exception:
+    MultiSpecSegformerService = None
+
+
+# ============================================================
+# Utils
+# ============================================================
+
+def now_ms():
+    return time.perf_counter() * 1000.0
+
+
+def mean(xs):
+    return float(np.mean(xs)) if xs else 0.0
+
+
+def p95(xs):
+    return float(np.percentile(xs, 95)) if xs else 0.0
+
+
+def maxv(xs):
+    return float(np.max(xs)) if xs else 0.0
+
+
+def last_mean(xs, n=30):
+    return float(np.mean(xs[-n:])) if xs else 0.0
+
+
+def last_max(xs, n=30):
+    return float(np.max(xs[-n:])) if xs else 0.0
+
+
+def load_json_if_exists(path):
+    if not path:
+        return None
+    with open(path, "r", encoding="utf-8") as f:
+        return json.load(f)
+
+
+def parse_float_list(s, expected=5, default=None):
+    if s is None:
+        return default
+
+    vals = [float(x.strip()) for x in str(s).split(",") if x.strip() != ""]
+
+    if len(vals) != expected:
+        raise ValueError(f"Esperado {expected} valores, veio {len(vals)}: {s}")
+
+    return np.array(vals, dtype=np.float32)
+
+
+def extract_mean_std_from_model_config(cfg):
+    if not isinstance(cfg, dict):
+        return None, None
+
+    mean_cfg = cfg.get("mean") or cfg.get("channel_mean") or cfg.get("norm_mean")
+    std_cfg = cfg.get("std") or cfg.get("channel_std") or cfg.get("norm_std")
+
+    norm = cfg.get("normalization") or cfg.get("norm") or {}
+
+    if mean_cfg is None and isinstance(norm, dict):
+        mean_cfg = norm.get("mean") or norm.get("channel_mean")
+
+    if std_cfg is None and isinstance(norm, dict):
+        std_cfg = norm.get("std") or norm.get("channel_std")
+
+    if mean_cfg is None or std_cfg is None:
+        return None, None
+
+    mean_arr = np.array(mean_cfg, dtype=np.float32)
+    std_arr = np.array(std_cfg, dtype=np.float32)
+
+    if mean_arr.size != 5 or std_arr.size != 5:
+        return None, None
+
+    return mean_arr, std_arr
+
+
+def tensor_stats(tensor):
+    names = ["R", "G", "B", "RE", "NIR"]
+    out = {}
+
+    for i, name in enumerate(names):
+        ch = tensor[i].astype(np.float32)
+        out[name] = {
+            "min": float(np.min(ch)),
+            "p01": float(np.percentile(ch, 1)),
+            "p50": float(np.percentile(ch, 50)),
+            "p99": float(np.percentile(ch, 99)),
+            "max": float(np.max(ch)),
+            "mean": float(np.mean(ch)),
+            "std": float(np.std(ch)),
+        }
+
+    return out
+
+
+def print_tensor_stats(label, tensor):
+    print("============================================")
+    print(f"[{label}] TENSOR")
+    print(f"shape={tensor.shape} dtype={tensor.dtype}")
+
+    stats = tensor_stats(tensor)
+    for ch, s in stats.items():
+        print(
+            f"{ch:>3} | "
+            f"min={s['min']:.4f} "
+            f"p01={s['p01']:.4f} "
+            f"p50={s['p50']:.4f} "
+            f"p99={s['p99']:.4f} "
+            f"max={s['max']:.4f} "
+            f"mean={s['mean']:.4f} "
+            f"std={s['std']:.4f}"
+        )
+
+    print("============================================")
+
+
+def to_u8_01(arr):
+    arr = np.asarray(arr, dtype=np.float32)
+    arr = np.nan_to_num(arr, nan=0.0, posinf=1.0, neginf=0.0)
+    arr = np.clip(arr, 0.0, 1.0)
+    return (arr * 255.0).astype(np.uint8)
+
+
+def make_panel(tensor):
+    rgb = np.stack([tensor[0], tensor[1], tensor[2]], axis=2)
+    rgb_bgr = cv2.cvtColor(to_u8_01(rgb), cv2.COLOR_RGB2BGR)
+
+    re_bgr = cv2.cvtColor(to_u8_01(tensor[3]), cv2.COLOR_GRAY2BGR)
+    nir_bgr = cv2.cvtColor(to_u8_01(tensor[4]), cv2.COLOR_GRAY2BGR)
+
+    false_rgb = np.stack(
+        [
+            tensor[4],  # visual R = NIR
+            tensor[3],  # visual G = RE
+            tensor[0],  # visual B = R real
+        ],
+        axis=2,
+    )
+    false_bgr = cv2.cvtColor(to_u8_01(false_rgb), cv2.COLOR_RGB2BGR)
+
+    def title(img, text):
+        h, w = img.shape[:2]
+        bar_h = 34
+        bar = np.zeros((bar_h, w, 3), dtype=np.uint8)
+        cv2.putText(
+            bar,
+            text,
+            (10, 24),
+            cv2.FONT_HERSHEY_SIMPLEX,
+            0.7,
+            (255, 255, 255),
+            2,
+            cv2.LINE_AA,
+        )
+        return np.vstack([bar, img])
+
+    rgb_bgr = title(rgb_bgr, "RGB")
+    re_bgr = title(re_bgr, "RE")
+    nir_bgr = title(nir_bgr, "NIR")
+    false_bgr = title(false_bgr, "Falso color NIR/RE/R")
+
+    top = np.hstack([rgb_bgr, re_bgr])
+    bottom = np.hstack([nir_bgr, false_bgr])
+
+    return np.vstack([top, bottom])
+
+
+# ============================================================
+# Benchmark processor
+# ============================================================
+
+class ScientificBenchmark:
+    def __init__(self, args):
+        self.args = args
+        self.target_size = (int(args.width), int(args.height))
+
+        self.model_cfg = load_json_if_exists(args.model_config_json)
+        mean_cfg, std_cfg = extract_mean_std_from_model_config(self.model_cfg)
+
+        if args.model_mean is not None:
+            self.mean = parse_float_list(args.model_mean, expected=5)
+        elif mean_cfg is not None:
+            self.mean = mean_cfg
+        else:
+            self.mean = np.array([0.5, 0.5, 0.5, 0.5, 0.5], dtype=np.float32)
+
+        if args.model_std is not None:
+            self.std = parse_float_list(args.model_std, expected=5)
+        elif std_cfg is not None:
+            self.std = std_cfg
+        else:
+            self.std = np.array([0.25, 0.25, 0.25, 0.25, 0.25], dtype=np.float32)
+
+        self.mean_chw = self.mean[:, None, None].astype(np.float32)
+        self.std_chw = self.std[:, None, None].astype(np.float32)
+
+        self.device = None
+        if torch is not None and torch.cuda.is_available():
+            self.device = torch.device("cuda")
+        elif torch is not None:
+            self.device = torch.device("cpu")
+
+        self.model_svc = None
+
+        if args.run_model:
+            if MultiSpecSegformerService is None:
+                raise RuntimeError("MultiSpecSegformerService não pôde ser importado.")
+
+            if not isinstance(self.model_cfg, dict):
+                raise RuntimeError("--run_model requer --model_config_json válido.")
+
+            self.model_svc = MultiSpecSegformerService(
+                model_config=self.model_cfg,
+                mostrar_log=print,
+            )
+
+            dummy = np.zeros((5, args.height, args.width), dtype=np.float32)
+
+            for _ in range(max(0, int(args.warmup_model))):
+                self.model_svc.infer_tensor_fast(dummy, keep_probs=False)
+
+            print(f"[BENCH] Warmup modelo concluído: {args.warmup_model}x")
+
+    def process_once(self, client):
+        """
+        Mede uma iteração completa do fluxo científico.
+        """
+        times = {
+            "capture_ms": 0.0,
+            "decode_ms": 0.0,
+            "controller_ms": 0.0,
+            "fuse_total_ms": 0.0,
+            "fuse_dark_ms": 0.0,
+            "fuse_radnorm_ms": 0.0,
+            "fuse_flat_ms": 0.0,
+            "fuse_prepare_ms": 0.0,
+            "fuse_warp_ms": 0.0,
+            "fuse_crop_resize_ms": 0.0,
+            "fuse_concat_ms": 0.0,
+            "model_norm_ms": 0.0,
+            "torch_ms": 0.0,
+            "infer_ms": 0.0,
+            "total_ms": 0.0,
+            "sync_dt_ms": 0.0,
+        }
+
+        t_total0 = now_ms()
+
+        # ========================================================
+        # 1. Captura RAW_BRUTO
+        # ========================================================
+        t0 = now_ms()
+        raw_frame, raw_meta = client.get_next_raw_frame(timeout=self.args.timeout)
+        times["capture_ms"] = now_ms() - t0
+
+        #cp = raw_meta.get("capture_perf", {})
+        #print(
+        #    "[CAP_ASYNC] "
+        #    f"get_wait={cp.get('async_get_wait_ms',0):.2f}ms "
+        #    f"age={cp.get('async_packet_age_ms',0):.2f}ms "
+        #    f"seq={cp.get('async_packet_seq')} "
+        #    f"thread_wait={cp.get('wait_total_ms',0):.2f}ms "
+        #    f"sleep={cp.get('sleep_ms',0):.2f}ms/{cp.get('sleep_count',0)} "
+        #    f"drain={cp.get('drain_total_ms',0):.2f}ms "
+        #    f"copy={cp.get('drain_frombuffer_copy_ms',0):.2f}ms "
+        #    f"status={cp.get('async_status',{})}"
+        #)
+
+        times["sync_dt_ms"] = float(raw_meta.get("sync_dt_ms", 0.0) or 0.0)
+
+        # ========================================================
+        # 2. Decode RAW10 packed -> float científico por câmera
+        # ========================================================
+        t0 = now_ms()
+        decoded = client.decode_stream_cameras(raw_frame, raw_meta)
+        times["decode_ms"] = now_ms() - t0
+
+        # ========================================================
+        # 3. RadiometricController update, se ativo
+        #    Isso NÃO é a radiometric_normalization do tensor.
+        #    É o controlador de exposição/ganho.
+        # ========================================================
+        t0 = now_ms()
+        client.update_radiometry(decoded, raw_meta)
+        times["controller_ms"] = now_ms() - t0
+
+        # ========================================================
+        # 4. Fusão científica no RawProcessorCore
+        #    dark + radnorm + flat + homografia + crop/resize + concat
+        # ========================================================
+        t0 = now_ms()
+        tensor = client.build_infer_tensor_from_decoded(
+            decoded=decoded,
+            meta=raw_meta,
+            channels_expected=5,
+            target_size=self.target_size,
+        )
+        times["fuse_total_ms"] = now_ms() - t0
+
+        # Pega detalhamento interno do core
+        try:
+            perf = (client.core.last_fusion_result or {}).get("perf", {}) or {}
+            times["fuse_dark_ms"] = float(perf.get("dark_ms", 0.0) or 0.0)
+            times["fuse_radnorm_ms"] = float(perf.get("radnorm_ms", 0.0) or 0.0)
+            times["fuse_flat_ms"] = float(perf.get("flat_ms", 0.0) or 0.0)
+            times["fuse_prepare_ms"] = float(perf.get("prepare_ms", 0.0) or 0.0)
+            times["fuse_warp_ms"] = float(perf.get("warp_total_ms", 0.0) or 0.0)
+            times["fuse_crop_resize_ms"] = float(perf.get("crop_resize_ms", 0.0) or 0.0)
+            times["fuse_concat_ms"] = float(perf.get("concat_ms", 0.0) or 0.0)
+        except Exception:
+            pass
+
+        # ========================================================
+        # 5. Normalização do modelo, opcional
+        # ========================================================
+        if self.args.simulate_model_norm:
+            t0 = now_ms()
+            tensor = (tensor - self.mean_chw) / np.maximum(self.std_chw, 1e-6)
+            tensor = np.ascontiguousarray(tensor, dtype=np.float32)
+            times["model_norm_ms"] = now_ms() - t0
+
+        # ========================================================
+        # 6. Transferência para torch/cuda, opcional
+        # ========================================================
+        if self.args.to_torch:
+            if torch is None:
+                raise RuntimeError("--to_torch requer torch instalado.")
+
+            t0 = now_ms()
+            x = torch.from_numpy(tensor).unsqueeze(0).to(self.device, non_blocking=True)
+
+            if self.device is not None and self.device.type == "cuda":
+                torch.cuda.synchronize()
+
+            times["torch_ms"] = now_ms() - t0
+
+        # ========================================================
+        # 7. Inferência real, opcional
+        # ========================================================
+        pred = None
+        if self.args.run_model:
+            t0 = now_ms()
+            pred = self.model_svc.infer_tensor_fast(tensor, keep_probs=False)
+            times["infer_ms"] = now_ms() - t0
+
+        times["total_ms"] = now_ms() - t_total0
+
+        return tensor, pred, raw_frame, raw_meta, decoded, times
+
+
+# ============================================================
+# Main
+# ============================================================
+
+def main():
+    ap = argparse.ArgumentParser(
+        description="Benchmark científico OAK-FCC-3 RAW_BRUTO -> tensor multiespectral final."
+    )
+
+    ap.add_argument(
+        "--module_params",
+        default=r"C:\ZendionInc\agrobot_base\Python\OAK\datasets\oak-fcc-3\calibration\module_params.json",
+        help="Caminho do module_params.json.",
+    )
+
+    ap.add_argument("--width", type=int, default=1024, help="Largura final do tensor.")
+    ap.add_argument("--height", type=int, default=640, help="Altura final do tensor.")
+    ap.add_argument("--fps", type=float, default=30.0)
+    ap.add_argument("--seconds", type=float, default=20.0)
+    ap.add_argument("--timeout", type=float, default=3.0)
+    ap.add_argument("--warmup", type=int, default=5)
+    ap.add_argument("--mx_id", default=None)
+    ap.add_argument("--sync_tolerance_ms", type=float, default=25.0)
+    ap.add_argument("--buffer_size", type=int, default=8)
+
+    ap.add_argument("--simulate_model_norm", action="store_true")
+    ap.add_argument("--model_mean", default=None)
+    ap.add_argument("--model_std", default=None)
+    ap.add_argument("--to_torch", action="store_true")
+
+    ap.add_argument("--run_model", action="store_true")
+    ap.add_argument("--model_config_json", default=None)
+    ap.add_argument("--warmup_model", type=int, default=3)
+
+    ap.add_argument("--save_debug", action="store_true")
+    ap.add_argument("--debug_dir", default="raw_bruto_scientific_benchmark")
+    ap.add_argument("--show", action="store_true")
+    ap.add_argument("--display_scale", type=float, default=0.65)
+
+    args = ap.parse_args()
+
+    bench = ScientificBenchmark(args)
+
+    client = OakFcc3Client(
+        width=args.width,
+        height=args.height,
+        fps=args.fps,
+        frame_type="RAW_BRUTO",
+        output_dtype="uint8",
+        capture_mode="TRIPLE",
+        raw_policy="require_triple",
+        module_calibration_json=args.module_params,
+        sync_tolerance_ms=args.sync_tolerance_ms,
+        buffer_size=args.buffer_size,
+        mx_id=args.mx_id,
+    )
+
+    client.core.warmup_numba_raw10_decode()
+
+    samples = {
+        "capture_ms": [],
+        "decode_ms": [],
+        "controller_ms": [],
+        "fuse_total_ms": [],
+        "fuse_dark_ms": [],
+        "fuse_radnorm_ms": [],
+        "fuse_flat_ms": [],
+        "fuse_prepare_ms": [],
+        "fuse_warp_ms": [],
+        "fuse_crop_resize_ms": [],
+        "fuse_concat_ms": [],
+        "model_norm_ms": [],
+        "torch_ms": [],
+        "infer_ms": [],
+        "total_ms": [],
+        "sync_dt_ms": [],
+    }
+
+    n_frames = 0
+    t_start = time.perf_counter()
+    t_last_log = t_start
+
+    last_tensor = None
+    last_meta = None
+
+    try:
+        client.start(print_debug=True)
+
+        # Warmup de câmera/controlador/filas
+        print(f"[BENCH] Warmup frames: {args.warmup}")
+        for _ in range(max(0, int(args.warmup))):
+            try:
+                bench.process_once(client)
+            except Exception as e:
+                print(f"[WARN] warmup falhou: {type(e).__name__}: {e}")
+            time.sleep(0.02)
+
+        print("============================================")
+        print("[BENCH] Iniciando benchmark científico RAW_BRUTO")
+        print(f"target tensor : (5,{args.height},{args.width})")
+        print(f"duration      : {args.seconds}s")
+        print("============================================")
+
+        while True:
+            now = time.perf_counter()
+            elapsed = now - t_start
+
+            if elapsed >= args.seconds:
+                break
+
+            try:
+                tensor, pred, raw_frame, raw_meta, decoded, times = bench.process_once(client)
+            except TimeoutError as e:
+                print(f"[TIMEOUT] {e}")
+                continue
+
+            n_frames += 1
+            last_tensor = tensor
+            last_meta = raw_meta
+
+            for k in samples:
+                samples[k].append(float(times.get(k, 0.0) or 0.0))
+
+            if now - t_last_log >= 1.0:
+                elapsed = now - t_start
+                fps = n_frames / max(elapsed, 1e-6)
+
+                print(
+                    "[RAW_SCI_PERF] "
+                    f"elapsed={elapsed:.1f}s "
+                    f"frames={n_frames} "
+                    f"fps={fps:.2f} "
+                    f"sync={last_mean(samples['sync_dt_ms']):.2f}ms "
+                    f"capture={last_mean(samples['capture_ms']):.2f}ms "
+                    f"decode={last_mean(samples['decode_ms']):.2f}ms "
+                    f"controller={last_mean(samples['controller_ms']):.2f}ms "
+                    f"fuse={last_mean(samples['fuse_total_ms']):.2f}ms "
+                    f"radnorm={last_mean(samples['fuse_radnorm_ms']):.2f}ms "
+                    f"flat={last_mean(samples['fuse_flat_ms']):.2f}ms "
+                    f"warp={last_mean(samples['fuse_warp_ms']):.2f}ms "
+                    f"crop_resize={last_mean(samples['fuse_crop_resize_ms']):.2f}ms "
+                    f"concat={last_mean(samples['fuse_concat_ms']):.2f}ms "
+                    f"model_norm={last_mean(samples['model_norm_ms']):.2f}ms "
+                    f"torch={last_mean(samples['torch_ms']):.2f}ms "
+                    f"infer={last_mean(samples['infer_ms']):.2f}ms "
+                    f"total={last_mean(samples['total_ms']):.2f}ms "
+                    f"tensor_shape={tuple(tensor.shape)}"
+                )
+
+                t_last_log = now
+
+        elapsed_total = time.perf_counter() - t_start
+        fps_total = n_frames / max(elapsed_total, 1e-6)
+
+        print("============================================")
+        print("RESULTADO FINAL RAW_BRUTO CIENTÍFICO")
+        print(f"elapsed : {elapsed_total:.2f}s")
+        print(f"frames  : {n_frames}")
+        print(f"fps     : {fps_total:.2f}")
+        print("--------------------------------------------")
+
+        def print_metric(name):
+            xs = samples[name]
+            print(
+                f"{name:18s} "
+                f"mean={mean(xs):8.2f}ms "
+                f"p95={p95(xs):8.2f}ms "
+                f"max={maxv(xs):8.2f}ms"
+            )
+
+        for name in [
+            "sync_dt_ms",
+            "capture_ms",
+            "decode_ms",
+            "controller_ms",
+            "fuse_total_ms",
+            "fuse_dark_ms",
+            "fuse_radnorm_ms",
+            "fuse_flat_ms",
+            "fuse_prepare_ms",
+            "fuse_warp_ms",
+            "fuse_crop_resize_ms",
+            "fuse_concat_ms",
+            "model_norm_ms",
+            "torch_ms",
+            "infer_ms",
+            "total_ms",
+        ]:
+            print_metric(name)
+
+        print("============================================")
+
+        if last_tensor is not None:
+            print_tensor_stats("LAST RAW_BRUTO SCI", last_tensor)
+
+            if args.save_debug:
+                debug_dir = Path(args.debug_dir)
+                debug_dir.mkdir(parents=True, exist_ok=True)
+
+                np.save(str(debug_dir / "last_tensor.npy"), last_tensor)
+
+                stats_path = debug_dir / "last_tensor_stats.json"
+                with open(stats_path, "w", encoding="utf-8") as f:
+                    json.dump(tensor_stats(last_tensor), f, indent=2, ensure_ascii=False)
+
+                panel = make_panel(last_tensor)
+                cv2.imwrite(str(debug_dir / "last_tensor_panel.png"), panel)
+
+                if last_meta is not None:
+                    with open(debug_dir / "last_meta.json", "w", encoding="utf-8") as f:
+                        json.dump(last_meta, f, indent=2, ensure_ascii=False)
+
+                print(f"[SAVE] Debug salvo em: {debug_dir}")
+
+            if args.show:
+                panel = make_panel(last_tensor)
+
+                if args.display_scale and abs(args.display_scale - 1.0) > 1e-6:
+                    new_w = max(1, int(panel.shape[1] * args.display_scale))
+                    new_h = max(1, int(panel.shape[0] * args.display_scale))
+                    panel = cv2.resize(panel, (new_w, new_h), interpolation=cv2.INTER_AREA)
+
+                cv2.imshow("RAW_BRUTO scientific tensor", panel)
+                print("[INFO] Pressione qualquer tecla para fechar.")
+                cv2.waitKey(0)
+                cv2.destroyAllWindows()
+
+    finally:
+        try:
+            client.stop()
+        except Exception:
+            pass
+
+
+if __name__ == "__main__":
+    main()

Python/OAK/datasets/oak-fcc-3/core/benchmark_preview.py

@@ -0,0 +1,645 @@
+import argparse
+import json
+import sys
+import time
+import threading
+from pathlib import Path
+
+import cv2
+import numpy as np
+
+# ============================================================
+# Ajuste de import local
+# ============================================================
+THIS_FILE = Path(__file__).resolve()
+WORKERS_DIR = THIS_FILE.parents[2]
+if str(WORKERS_DIR) not in sys.path:
+    sys.path.insert(0, str(WORKERS_DIR))
+
+from camera_worker.oak_fcc3_core.oak_fcc3_client import OakFcc3Client
+
+try:
+    from camera_worker.oak_fcc3_core.segformer_service import MultiSpecSegformerService
+except Exception:
+    MultiSpecSegformerService = None
+
+
+# ============================================================
+# FPS / utilidades
+# ============================================================
+class FpsMeter:
+    def __init__(self, alpha=0.15):
+        self.alpha = float(alpha)
+        self.last_ts = None
+        self.fps = 0.0
+
+    def tick(self):
+        now = time.perf_counter()
+        if self.last_ts is not None:
+            dt = now - self.last_ts
+            if dt > 1e-9:
+                inst = 1.0 / dt
+                self.fps = inst if self.fps <= 0 else (1.0 - self.alpha) * self.fps + self.alpha * inst
+        self.last_ts = now
+        return self.fps
+
+
+def load_json_if_exists(path):
+    if not path:
+        return None
+    with open(path, "r", encoding="utf-8") as f:
+        return json.load(f)
+
+
+def init_model_service(args):
+    """
+    Mesmo contrato do benchmark científico:
+      --run_model exige --model_config_json
+      MultiSpecSegformerService(model_config=cfg).infer_tensor_fast(tensor, keep_probs=False)
+    """
+    if not args.run_model:
+        return None
+
+    if MultiSpecSegformerService is None:
+        raise RuntimeError("MultiSpecSegformerService não pôde ser importado.")
+
+    model_cfg = load_json_if_exists(args.model_config_json)
+    if not isinstance(model_cfg, dict):
+        raise RuntimeError("--run_model requer --model_config_json válido.")
+
+    svc = MultiSpecSegformerService(
+        model_config=model_cfg,
+        mostrar_log=print,
+    )
+
+    dummy = np.zeros((5, int(args.height), int(args.width)), dtype=np.float32)
+    for _ in range(max(0, int(args.warmup_model))):
+        svc.infer_tensor_fast(dummy, keep_probs=False)
+
+    print(f"[MODEL] Warmup concluído: {args.warmup_model}x")
+    return svc
+
+
+def to_u8_01(arr, auto_level=False):
+    arr = np.asarray(arr, dtype=np.float32)
+    arr = np.nan_to_num(arr, nan=0.0, posinf=1.0, neginf=0.0)
+
+    if auto_level:
+        p1 = float(np.percentile(arr, 1))
+        p99 = float(np.percentile(arr, 99))
+        den = max(p99 - p1, 1e-6)
+        arr = (arr - p1) / den
+
+    arr = np.clip(arr, 0.0, 1.0)
+    return (arr * 255.0).astype(np.uint8)
+
+
+def ensure_bgr(img, auto_level=False):
+    img = np.asarray(img)
+
+    if img.ndim == 2:
+        g = to_u8_01(img, auto_level=auto_level)
+        return cv2.cvtColor(g, cv2.COLOR_GRAY2BGR)
+
+    if img.ndim == 3 and img.shape[2] == 3:
+        u8 = to_u8_01(img, auto_level=auto_level)
+        # decoded/tensor RGB vem em RGB; OpenCV mostra BGR
+        return cv2.cvtColor(u8, cv2.COLOR_RGB2BGR)
+
+    raise RuntimeError(f"Imagem inválida para visualização: shape={img.shape}")
+
+
+def tensor_rgb_to_bgr(tensor, auto_level=False):
+    rgb = np.stack([tensor[0], tensor[1], tensor[2]], axis=2)
+    return ensure_bgr(rgb, auto_level=auto_level)
+
+
+def tensor_channel_to_bgr(tensor, idx, auto_level=False):
+    return ensure_bgr(tensor[idx], auto_level=auto_level)
+
+
+def add_title(img, title, color=(255, 255, 255)):
+    out = img.copy()
+    h, w = out.shape[:2]
+    bar_h = 34
+    bar = np.zeros((bar_h, w, 3), dtype=np.uint8)
+    cv2.putText(bar, str(title), (10, 23), cv2.FONT_HERSHEY_SIMPLEX, 0.62, color, 2, cv2.LINE_AA)
+    return np.vstack([bar, out])
+
+
+def add_hud(panel, lines):
+    out = panel.copy()
+    x, y = 12, 45
+    for line in lines:
+        cv2.putText(out, line, (x, y), cv2.FONT_HERSHEY_SIMPLEX, 0.62, (0, 255, 255), 2, cv2.LINE_AA)
+        y += 24
+    return out
+
+
+def resize_tile(img, tile_w, tile_h):
+    return cv2.resize(img, (int(tile_w), int(tile_h)), interpolation=cv2.INTER_AREA)
+
+
+def get_cam_by_role(decoded, role):
+    role = str(role).lower()
+    for cam_id, item in decoded.items():
+        r = str(item.get("role") or item.get("meta", {}).get("role") or "").lower()
+        if r == role:
+            return cam_id
+    return None
+
+
+def decoded_raw_tiles(decoded, tile_w, tile_h, auto_level=False):
+    """
+    Retorna tiles RAW/decoded para RGB, RE, NIR antes da fusão final.
+    Aqui 'RAW_BRUTO' significa o conteúdo decodificado vindo das câmeras, ainda no espaço nativo.
+    """
+    tiles = {}
+
+    rgb_id = get_cam_by_role(decoded, "rgb")
+    re_id = get_cam_by_role(decoded, "re")
+    nir_id = get_cam_by_role(decoded, "nir")
+
+    if rgb_id is not None:
+        img = decoded[rgb_id]["image"]
+        tiles["rgb"] = resize_tile(ensure_bgr(img, auto_level=auto_level), tile_w, tile_h)
+    else:
+        tiles["rgb"] = np.zeros((tile_h, tile_w, 3), dtype=np.uint8)
+
+    if re_id is not None:
+        img = decoded[re_id]["image"]
+        tiles["re"] = resize_tile(ensure_bgr(img, auto_level=auto_level), tile_w, tile_h)
+    else:
+        tiles["re"] = np.zeros((tile_h, tile_w, 3), dtype=np.uint8)
+
+    if nir_id is not None:
+        img = decoded[nir_id]["image"]
+        tiles["nir"] = resize_tile(ensure_bgr(img, auto_level=auto_level), tile_w, tile_h)
+    else:
+        tiles["nir"] = np.zeros((tile_h, tile_w, 3), dtype=np.uint8)
+
+    return tiles
+
+
+def tensor_tiles(tensor, tile_w, tile_h, auto_level=False):
+    return {
+        "rgb": resize_tile(tensor_rgb_to_bgr(tensor, auto_level=auto_level), tile_w, tile_h),
+        "re": resize_tile(tensor_channel_to_bgr(tensor, 3, auto_level=auto_level), tile_w, tile_h),
+        "nir": resize_tile(tensor_channel_to_bgr(tensor, 4, auto_level=auto_level), tile_w, tile_h),
+    }
+
+
+def colorize_label_map(label_map, num_classes=None):
+    label = np.asarray(label_map)
+    if label.ndim == 3:
+        label = np.argmax(label, axis=0)
+    label = label.astype(np.int32)
+
+    if num_classes is None:
+        num_classes = int(max(1, label.max() + 1))
+
+    # Paleta simples e estável. BGR.
+    palette = np.array([
+        [40, 40, 40],
+        [60, 180, 60],
+        [60, 60, 220],
+        [220, 180, 60],
+        [180, 60, 180],
+        [180, 180, 60],
+        [60, 180, 180],
+        [220, 220, 220],
+    ], dtype=np.uint8)
+
+    out = palette[label % len(palette)]
+    return out
+
+
+def try_extract_prediction_tiles(pred, target_w, target_h):
+    """
+    Tentativa genérica. Adapte aqui se o benchmark tiver nomes específicos das cabeças.
+    Retorna até 3 tiles BGR: semântica/head0/head1.
+    """
+    if pred is None:
+        blank = np.zeros((target_h, target_w, 3), dtype=np.uint8)
+        return [blank, blank.copy(), blank.copy()], ["Pred vazio", "Head 1", "Head 2"]
+
+    candidates = []
+    names = []
+
+    if isinstance(pred, dict):
+        # nomes comuns
+        for key in ("mask", "pred_mask", "class_map", "semantic", "semantic_mask", "segmentation"):
+            if key in pred:
+                candidates.append(pred[key])
+                names.append(key)
+
+        heads = pred.get("heads") or pred.get("head_outputs") or pred.get("predictions")
+        if isinstance(heads, dict):
+            for k, v in heads.items():
+                candidates.append(v)
+                names.append(str(k))
+        elif isinstance(heads, (list, tuple)):
+            for i, v in enumerate(heads):
+                candidates.append(v)
+                names.append(f"head_{i}")
+    else:
+        candidates.append(pred)
+        names.append("prediction")
+
+    tiles = []
+    out_names = []
+
+    for name, arr in zip(names, candidates):
+        arr = np.asarray(arr)
+
+        # remove batch se existir
+        if arr.ndim == 4 and arr.shape[0] == 1:
+            arr = arr[0]
+
+        if arr.ndim == 3:
+            # CHW logits/probs ou HWC RGB/probs
+            if arr.shape[0] <= 32:
+                vis = colorize_label_map(np.argmax(arr, axis=0), num_classes=arr.shape[0])
+            elif arr.shape[2] in (1, 3):
+                vis = ensure_bgr(arr[:, :, 0] if arr.shape[2] == 1 else arr, auto_level=True)
+            else:
+                vis = ensure_bgr(np.max(arr, axis=2), auto_level=True)
+        elif arr.ndim == 2:
+            # se parecer label map, colore; se parecer float, cinza auto-level
+            if np.issubdtype(arr.dtype, np.integer) and int(np.max(arr)) <= 64:
+                vis = colorize_label_map(arr)
+            else:
+                vis = ensure_bgr(arr, auto_level=True)
+        elif arr.ndim == 1:
+            # vetor de classe/score: desenha texto
+            vis = np.zeros((target_h, target_w, 3), dtype=np.uint8)
+            txt = np.array2string(arr[:8], precision=2, separator=", ")
+            cv2.putText(vis, txt[:80], (10, target_h // 2), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (255, 255, 255), 1, cv2.LINE_AA)
+        else:
+            continue
+
+        vis = resize_tile(vis, target_w, target_h)
+        tiles.append(vis)
+        out_names.append(name)
+
+        if len(tiles) >= 3:
+            break
+
+    while len(tiles) < 3:
+        tiles.append(np.zeros((target_h, target_w, 3), dtype=np.uint8))
+        out_names.append(f"pred_{len(tiles)}")
+
+    return tiles[:3], out_names[:3]
+
+
+def try_run_model(model_svc, tensor):
+    """
+    Inferência real, igual ao benchmark científico.
+    Mantém esta função isolada para adaptar fácil caso o retorno do modelo mude.
+    """
+    if model_svc is None:
+        return None, "model_svc_none"
+
+    if hasattr(model_svc, "infer_tensor_fast"):
+        return model_svc.infer_tensor_fast(tensor, keep_probs=False), None
+
+    if hasattr(model_svc, "infer"):
+        return model_svc.infer(tensor), None
+
+    return None, "model_svc_sem_infer"
+
+
+def build_grid(raw_tiles, final_tiles, pred_tiles=None, pred_names=None, tile_w=420, tile_h=260):
+    rows = []
+    row_defs = [
+        ("RGB", "rgb"),
+        ("RE", "re"),
+        ("NIR", "nir"),
+    ]
+
+    for i, (label, key) in enumerate(row_defs):
+        left = add_title(raw_tiles[key], f"RAW_BRUTO decoded {label}")
+        mid = add_title(final_tiles[key], f"Tensor final {label}")
+
+        cells = [left, mid]
+
+        if pred_tiles is not None:
+            name = pred_names[i] if pred_names and i < len(pred_names) else f"Pred {i}"
+            cells.append(add_title(pred_tiles[i], name))
+
+        # iguala altura após título
+        h_min = min(c.shape[0] for c in cells)
+        norm = [cv2.resize(c, (tile_w, h_min), interpolation=cv2.INTER_AREA) if c.shape[1] != tile_w or c.shape[0] != h_min else c for c in cells]
+        rows.append(np.hstack(norm))
+
+    return np.vstack(rows)
+
+
+def get_core_perf(client):
+    for attr in ("raw_processor", "processor", "core", "raw_processor_core"):
+        obj = getattr(client, attr, None)
+        if obj is not None and getattr(obj, "last_fusion_result", None) is not None:
+            return obj.last_fusion_result.get("perf", {}) or {}
+    return {}
+
+
+
+# ============================================================
+# Estado compartilhado / workers assíncronos
+# ============================================================
+class SharedState:
+    def __init__(self):
+        self.lock = threading.Lock()
+        self.running = True
+        self.latest_decoded = None
+        self.latest_meta = None
+        self.latest_tensor = None
+        self.latest_perf = {}
+        self.latest_pred = None
+        self.latest_model_warn = None
+        self.latest_model_ms = 0.0
+        self.latest_error = None
+        self.tensor_fps = FpsMeter()
+        self.model_fps = FpsMeter()
+        self.preview_fps = FpsMeter()
+        self.tensor_seq = 0
+        self.model_seq = 0
+
+    def stop(self):
+        with self.lock:
+            self.running = False
+
+    def is_running(self):
+        with self.lock:
+            return bool(self.running)
+
+
+def tensor_worker(client, state, args):
+    """
+    Roda no talo: captura RAW_BRUTO, monta tensor final e atualiza cache.
+    Não depende do FPS da janela.
+    """
+    while state.is_running():
+        try:
+            frame, meta, decoded = client.get_next_decoded(timeout=args.timeout)
+            if not isinstance(frame, dict):
+                raise RuntimeError(f"RAW_BRUTO esperado como dict. Veio {type(frame)}")
+
+            tensor = client.build_infer_tensor_from_decoded(
+                decoded=decoded,
+                meta=meta,
+                channels_expected=5,
+                target_size=(args.width, args.height),
+            )
+            tensor = np.ascontiguousarray(tensor.astype(np.float32, copy=False))
+            perf = get_core_perf(client)
+            fps = state.tensor_fps.tick()
+
+            with state.lock:
+                state.latest_decoded = decoded
+                state.latest_meta = meta
+                state.latest_tensor = tensor
+                state.latest_perf = dict(perf or {})
+                state.tensor_seq += 1
+                state.latest_error = None
+
+        except Exception as e:
+            with state.lock:
+                state.latest_error = f"tensor_worker: {type(e).__name__}: {e}"
+            time.sleep(0.02)
+
+
+def model_worker(model_svc, state, args):
+    """
+    Opcional: roda inferência no último tensor disponível.
+    Não bloqueia o worker de tensor nem a janela.
+    """
+    last_seq = -1
+
+    while state.is_running():
+        with state.lock:
+            tensor = None if state.latest_tensor is None else state.latest_tensor.copy()
+            seq = state.tensor_seq
+
+        if tensor is None or seq == last_seq:
+            time.sleep(0.005)
+            continue
+
+        last_seq = seq
+
+        try:
+            t0_model = time.perf_counter()
+            pred, warn = try_run_model(model_svc, tensor)
+            model_ms = (time.perf_counter() - t0_model) * 1000.0
+            if warn is None:
+                state.model_fps.tick()
+
+            with state.lock:
+                state.latest_pred = pred
+                state.latest_model_warn = warn
+                state.latest_model_ms = float(model_ms)
+                state.model_seq += 1
+
+        except Exception as e:
+            with state.lock:
+                state.latest_model_warn = f"model_worker: {type(e).__name__}: {e}"
+            time.sleep(0.02)
+
+
+def snapshot_state(state):
+    """
+    Copia referências do cache para desenhar. A janela só roda na cadência do preview.
+    """
+    with state.lock:
+        return {
+            "decoded": state.latest_decoded,
+            "meta": state.latest_meta,
+            "tensor": state.latest_tensor,
+            "perf": dict(state.latest_perf or {}),
+            "pred": state.latest_pred,
+            "model_warn": state.latest_model_warn,
+            "model_ms": state.latest_model_ms,
+            "error": state.latest_error,
+            "tensor_fps": state.tensor_fps.fps,
+            "model_fps": state.model_fps.fps,
+            "tensor_seq": state.tensor_seq,
+            "model_seq": state.model_seq,
+        }
+
+
+# ============================================================
+# Main loop assíncrono
+# ============================================================
+def main():
+    ap = argparse.ArgumentParser(description="Preview assíncrono RAW_BRUTO decoded vs tensor final multispectral.")
+    ap.add_argument("--module_params", default=r"C:\ZendionInc\agrobot_base\Python\OAK\datasets\oak-fcc-3\calibration\module_params.json")
+    ap.add_argument("--width", type=int, default=1024, help="Largura final do tensor.")
+    ap.add_argument("--height", type=int, default=640, help="Altura final do tensor.")
+    ap.add_argument("--fps", type=float, default=40.0, help="FPS alvo da câmera.")
+    ap.add_argument("--preview_fps", type=float, default=5.0, help="FPS da janela OpenCV apenas.")
+    ap.add_argument("--timeout", type=float, default=2.0)
+    ap.add_argument("--warmup", type=int, default=5)
+    ap.add_argument("--mx_id", default=None)
+    ap.add_argument("--display_scale", type=float, default=0.75)
+    ap.add_argument("--tile_w", type=int, default=420)
+    ap.add_argument("--tile_h", type=int, default=260)
+    ap.add_argument("--auto_level_raw", action="store_true")
+    ap.add_argument("--auto_level_tensor", action="store_true")
+    ap.add_argument("--run_model", action="store_true", help="Roda inferência em thread separada usando o último tensor cacheado.")
+    ap.add_argument("--model_config_json", default=None, help="JSON de configuração do modelo SegFormer, igual ao benchmark.")
+    ap.add_argument("--warmup_model", type=int, default=3, help="Número de inferências dummy para aquecer o modelo.")
+    ap.add_argument("--save_last", default=None)
+    args = ap.parse_args()
+
+    client = OakFcc3Client(
+        width=args.width,
+        height=args.height,
+        fps=args.fps,
+        frame_type="RAW_BRUTO",
+        capture_mode="TRIPLE",
+        raw_policy="require_triple",
+        module_calibration_json=args.module_params,
+        mx_id=args.mx_id,
+    )
+
+    model_svc = init_model_service(args) if args.run_model else None
+
+    state = SharedState()
+    last_panel = None
+    last_warn_ts = 0.0
+    last_seq_drawn = -1
+
+    try:
+        client.start(print_debug=True)
+
+        for _ in range(max(0, int(args.warmup))):
+            try:
+                client.get_next_decoded(timeout=args.timeout)
+            except Exception:
+                pass
+            time.sleep(0.03)
+
+        tw = threading.Thread(target=tensor_worker, args=(client, state, args), daemon=True)
+        tw.start()
+
+        mw = None
+        if args.run_model:
+            mw = threading.Thread(target=model_worker, args=(model_svc, state, args), daemon=True)
+            mw.start()
+
+        min_period = 1.0 / max(float(args.preview_fps), 0.1)
+        next_draw_ts = 0.0
+
+        print("[INFO] Preview assíncrono iniciado. Pressione Q ou ESC para sair.")
+        print("[INFO] Tensor FPS = geração real do tensor. Preview FPS = janela. Model FPS = inferência, se habilitada.")
+
+        while True:
+            now = time.perf_counter()
+            if now < next_draw_ts:
+                time.sleep(min(0.005, next_draw_ts - now))
+                key = cv2.waitKey(1) & 0xFF
+                if key in (27, ord('q'), ord('Q')):
+                    break
+                continue
+
+            next_draw_ts = now + min_period
+            snap = snapshot_state(state)
+
+            if snap["tensor"] is None or snap["decoded"] is None:
+                blank = np.zeros((360, 900, 3), dtype=np.uint8)
+                cv2.putText(blank, "Aguardando primeiro tensor...", (30, 180), cv2.FONT_HERSHEY_SIMPLEX, 0.9, (255,255,255), 2, cv2.LINE_AA)
+                cv2.imshow("OAK RAW_BRUTO vs Tensor Final", blank)
+                key = cv2.waitKey(1) & 0xFF
+                if key in (27, ord('q'), ord('Q')):
+                    break
+                continue
+
+            # Desenha só usando o cache. Não captura nem monta tensor aqui.
+            t_draw0 = time.perf_counter()
+            raw_tiles = decoded_raw_tiles(
+                snap["decoded"],
+                tile_w=args.tile_w,
+                tile_h=args.tile_h,
+                auto_level=args.auto_level_raw,
+            )
+            final_tiles = tensor_tiles(
+                snap["tensor"],
+                tile_w=args.tile_w,
+                tile_h=args.tile_h,
+                auto_level=args.auto_level_tensor,
+            )
+
+            pred_tiles = None
+            pred_names = None
+            if args.run_model:
+                pred_tiles, pred_names = try_extract_prediction_tiles(
+                    snap["pred"],
+                    target_w=args.tile_w,
+                    target_h=args.tile_h,
+                )
+                if snap["model_warn"]:
+                    tnow = time.time()
+                    if tnow - last_warn_ts > 2.0:
+                        print(f"[WARN][MODEL] {snap['model_warn']}")
+                        last_warn_ts = tnow
+
+            panel = build_grid(
+                raw_tiles=raw_tiles,
+                final_tiles=final_tiles,
+                pred_tiles=pred_tiles,
+                pred_names=pred_names,
+                tile_w=args.tile_w,
+                tile_h=args.tile_h,
+            )
+
+            preview_fps = state.preview_fps.tick()
+            draw_ms = (time.perf_counter() - t_draw0) * 1000.0
+            perf = snap["perf"]
+            meta = snap["meta"] or {}
+            tensor_seq = int(snap["tensor_seq"])
+            dropped_for_preview = max(0, tensor_seq - last_seq_drawn - 1) if last_seq_drawn >= 0 else 0
+            last_seq_drawn = tensor_seq
+
+            hud = [
+                f"Preview FPS: {preview_fps:.1f} | Tensor FPS: {snap['tensor_fps']:.1f} | Model FPS: {snap['model_fps']:.1f} | infer={snap.get('model_ms', 0.0):.1f}ms",
+                f"draw={draw_ms:.1f}ms flat={perf.get('flat_ms', 0):.1f} warp={perf.get('warp_total_ms', 0):.1f} crop={perf.get('crop_resize_ms', 0):.1f} fuse={perf.get('total_ms', 0):.1f}",
+                f"seq={tensor_seq} skipped_preview={dropped_for_preview} frame_type={meta.get('frame_type')} run_model={args.run_model}",
+            ]
+            if snap["error"]:
+                hud.append(str(snap["error"])[:120])
+
+            panel = add_hud(panel, hud)
+            last_panel = panel
+
+            disp = panel
+            if args.display_scale and abs(args.display_scale - 1.0) > 1e-6:
+                disp = cv2.resize(
+                    disp,
+                    (max(1, int(disp.shape[1] * args.display_scale)), max(1, int(disp.shape[0] * args.display_scale))),
+                    interpolation=cv2.INTER_AREA,
+                )
+
+            cv2.imshow("OAK RAW_BRUTO vs Tensor Final", disp)
+            key = cv2.waitKey(1) & 0xFF
+            if key in (27, ord('q'), ord('Q')):
+                break
+
+    finally:
+        state.stop()
+        time.sleep(0.05)
+
+        if args.save_last and last_panel is not None:
+            out_path = Path(args.save_last)
+            out_path.parent.mkdir(parents=True, exist_ok=True)
+            cv2.imwrite(str(out_path), last_panel)
+            print(f"[SAVE] {out_path}")
+
+        try:
+            client.stop()
+        except Exception:
+            pass
+        cv2.destroyAllWindows()
+
+
+if __name__ == "__main__":
+    main()

Python/OAK/datasets/oak-fcc-3/core/oak_fcc3_client.py

@@ -23,6 +23,7 @@ class OakFcc3Client:
         module_calibration_json=None,
         sync_mode="best",
         sync_tolerance_ms=25.0,
+        mx_id=None,
         **kwargs,
     ):
         self.width = width
@@ -37,6 +38,8 @@ class OakFcc3Client:
         self.module_params = self._load_module_params(module_calibration_json)
         self.fusion_config = self.module_params.get("fusion_config", {}) or {}
 
+        self.mx_id = str(mx_id) if mx_id else None
+
         self.svc = OakFcc3Service(
             timeout=10,
             fps=fps,
@@ -48,6 +51,8 @@ class OakFcc3Client:
             raw_policy=raw_policy,
             sync_mode=sync_mode,
             sync_tolerance_ms=sync_tolerance_ms,
+            mx_id=self.mx_id,
+            module_calibration_json=module_calibration_json,
             **kwargs,
         )
 
@@ -152,6 +157,11 @@ class OakFcc3Client:
             capture_mode=self.capture_mode,
         )
 
+        try:
+            self.mx_id = self.svc.manager.mx_id
+        except Exception:
+            pass
+
         applied = self.apply_module_camera_settings()
 
         if print_debug:
@@ -181,39 +191,56 @@ class OakFcc3Client:
     def get_next_decoded(self, timeout=2.0):
         raw_frame, raw_meta = self.get_next_raw_frame(timeout=timeout)
 
-        decoded = self.decode_stream_cameras(raw_frame, raw_meta)
-
-        self.update_radiometry(decoded, raw_meta)
-
         frame_type = str(raw_meta.get("frame_type", self.frame_type)).upper()
-
         meta = dict(raw_meta)
 
         if frame_type == "RAW_BRUTO":
+            decoded = self.decode_stream_cameras(raw_frame, raw_meta)
+
+            self.update_radiometry(decoded, raw_meta)
+
             frame = raw_frame
+            return frame, meta, decoded
 
         elif frame_type == "RGB":
+            decoded = self.decode_stream_cameras(raw_frame, raw_meta)
+
+            self.update_radiometry(decoded, raw_meta)
+
             frame = self.build_rgb_tensor(decoded)
             meta["output_layout"] = "CHW"
             meta["channels"] = ["R", "G", "B"]
             meta["shape"] = list(frame.shape)
             meta["dtype"] = str(frame.dtype)
 
+            return frame, meta, decoded
+
         elif frame_type == "MULTISPEC":
-            frame = self.build_multispec_tensor(decoded, meta=raw_meta)
+            decoded = self.decode_oak_aligned_multispec(raw_frame, raw_meta)
+
+            self.update_radiometry(decoded, raw_meta)
+
+            # Versão inicial segura:
+            # não chama core.fuse_multispec_cameras(), porque ali teria homografia de novo.
+            frame = self.build_multispec_tensor_from_oak_aligned(decoded, meta=raw_meta)
+
             meta["output_layout"] = "CHW"
             meta["channels"] = ["R", "G", "B", "RE", "NIR"]
             meta["shape"] = list(frame.shape)
             meta["dtype"] = str(frame.dtype)
+            meta["aligned_by_oak"] = True
+            meta["geometry_stage"] = "oak"
+
+            return frame, meta, decoded
 
         elif frame_type == "PREVIEW":
+            decoded = self.decode_stream_cameras(raw_frame, raw_meta)
             frame = raw_frame
+            return frame, meta, decoded
 
         else:
             raise RuntimeError(f"frame_type não suportado: {frame_type}")
 
-        return frame, meta, decoded
-
     def get_next_tensor_preview(self, timeout=2.0):
         frame, meta, decoded = self.get_next_decoded(timeout=timeout)
 
@@ -551,3 +578,95 @@ class OakFcc3Client:
     def _gray01_to_bgr(gray01):
         g = np.clip(gray01 * 255.0, 0, 255).astype(np.uint8)
         return cv2.cvtColor(g, cv2.COLOR_GRAY2BGR)
+
+
+
+    def decode_oak_aligned_multispec(self, frame, meta):
+        """
+        Decodifica frames já alinhados pela OAK.
+
+        Entrada esperada:
+            frame = {
+                "CAM_A": BGR uint8 HWC,
+                "CAM_B": GRAY uint8 HW,
+                "CAM_C": GRAY uint8 HW,
+            }
+
+        Saída:
+            decoded por role, em float32 0..1.
+        """
+        if not isinstance(frame, dict):
+            raise RuntimeError("MULTISPEC alinhado esperado como dict de câmeras.")
+
+        decoded = {}
+        camera_info = meta.get("camera_info", {}) or {}
+
+        for cam_id, img in frame.items():
+            info = camera_info.get(cam_id, {}) or {}
+            role = str(info.get("role", "")).lower()
+
+            if not role:
+                role = str(self.svc.manager.roles.get(cam_id, cam_id)).lower()
+
+            img01 = self._frame_to_float01(cam_id, img, role)
+
+            decoded[cam_id] = {
+                "name": role.upper(),
+                "role": role,
+                "image": img01,
+                "meta": {
+                    **info,
+                    "cam_id": cam_id,
+                    "role": role,
+                    "aligned_by_oak": True,
+                    "geometry_stage": "oak",
+                    "homography_applied": role in ("re", "nir"),
+                    "crop_resize_applied": True,
+                    "shape": list(img.shape),
+                    "dtype": str(img.dtype),
+                },
+            }
+
+        return decoded
+
+    def build_multispec_tensor_from_oak_aligned(self, decoded, meta=None):
+        """
+        Monta CHW [R,G,B,RE,NIR] sem reaplicar homografia.
+        """
+        _, rgb_item = self._find_decoded_by_role(decoded, "rgb")
+        _, re_item = self._find_decoded_by_role(decoded, "re")
+        _, nir_item = self._find_decoded_by_role(decoded, "nir")
+
+        rgb = rgb_item["image"].astype(np.float32, copy=False)
+        re = re_item["image"].astype(np.float32, copy=False)
+        nir = nir_item["image"].astype(np.float32, copy=False)
+
+        if rgb.ndim != 3 or rgb.shape[2] != 3:
+            raise RuntimeError(f"RGB alinhado inválido: shape={rgb.shape}")
+
+        h, w = rgb.shape[:2]
+
+        if re.ndim == 3:
+            re = re[:, :, 0]
+
+        if nir.ndim == 3:
+            nir = nir[:, :, 0]
+
+        if re.shape[:2] != (h, w):
+            re = cv2.resize(re, (w, h), interpolation=cv2.INTER_LINEAR)
+
+        if nir.shape[:2] != (h, w):
+            nir = cv2.resize(nir, (w, h), interpolation=cv2.INTER_LINEAR)
+
+        tensor = np.stack(
+            [
+                rgb[:, :, 0],  # R
+                rgb[:, :, 1],  # G
+                rgb[:, :, 2],  # B
+                re,
+                nir,
+            ],
+            axis=0,
+        ).astype(np.float32, copy=False)
+
+        return np.ascontiguousarray(tensor)

Python/OAK/datasets/oak-fcc-3/core/oak_fcc3_service.py

@@ -45,6 +45,7 @@ class OakFcc3Service:
 
     def get_config(self):
         return {
+            "mx_id": self.manager.mx_id,
             "ok": True,
             "fps": self.manager.fps,
             "width": self.manager.width,

Python/OAK/datasets/oak-fcc-3/core/segformer_service.py

@@ -0,0 +1,915 @@
+# camera_worker/multispec_segformer_service.py
+# -*- coding: utf-8 -*-
+
+from __future__ import annotations
+
+import copy
+import json
+import time
+from pathlib import Path
+from typing import Dict, Optional, Sequence, Tuple
+
+import cv2
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from transformers import SegformerConfig, SegformerForSemanticSegmentation
+
+
+DEFAULT_HEADS = {
+    "semantic": {
+        "enabled": True,
+        "type": "multiclass",
+        "num_classes": 3,
+        "classes": {"chao": 0, "cana": 1, "erva": 2},
+        "ignore_index": 255,
+    },
+    "vegetation": {
+        "enabled": True,
+        "type": "binary",
+        "num_classes": 2,
+        "classes": {"background": 0, "vegetation": 1},
+        "ignore_index": 255,
+    },
+    "cana": {
+        "enabled": True,
+        "type": "binary",
+        "num_classes": 2,
+        "classes": {"not_cana": 0, "cana": 1},
+        "ignore_index": 255,
+    },
+}
+
+SEMANTIC_COLORS_RGB = {
+    0: (85, 85, 85),      # chao
+    1: (0, 190, 0),       # cana
+    2: (230, 55, 55),     # erva
+}
+
+BINARY_COLORS_RGB = {
+    0: (30, 30, 30),
+    1: (0, 220, 80),
+}
+
+CANA_COLORS_RGB = {
+    0: (30, 30, 30),
+    1: (40, 210, 255),
+}
+
+TARGET_COLORS_RGB = {
+    0: (30, 30, 30),
+    1: (255, 70, 30),
+}
+
+
+def load_json(path: str | Path) -> dict:
+    with open(path, "r", encoding="utf-8") as f:
+        return json.load(f)
+
+
+def merge_dict(dst: dict, src: dict) -> dict:
+    out = copy.deepcopy(dst)
+
+    def rec(a, b):
+        for k, v in b.items():
+            if isinstance(v, dict) and isinstance(a.get(k), dict):
+                rec(a[k], v)
+            else:
+                a[k] = v
+
+    if isinstance(src, dict):
+        rec(out, src)
+
+    return out
+
+
+def build_heads_config(config: dict, ignore_index: int = 255) -> Dict[str, dict]:
+    cfg = merge_dict(DEFAULT_HEADS, config.get("heads", {}) or {})
+    active = {}
+
+    for name, hcfg in cfg.items():
+        if not bool(hcfg.get("enabled", True)):
+            continue
+
+        hcfg.setdefault("ignore_index", ignore_index)
+        hcfg["ignore_index"] = int(hcfg.get("ignore_index", ignore_index))
+        hcfg["num_classes"] = int(hcfg.get("num_classes", 2))
+        active[name] = hcfg
+
+    for required in ("semantic", "vegetation", "cana"):
+        if required not in active:
+            raise RuntimeError(f"Head obrigatória ausente no config: {required}")
+
+    return active
+
+
+def patch_segformer_encoder_input_channels(segformer_encoder: nn.Module, in_ch: int):
+    if in_ch == 3:
+        return segformer_encoder
+
+    proj = segformer_encoder.encoder.patch_embeddings[0].proj
+
+    if proj.in_channels == in_ch:
+        return segformer_encoder
+
+    old_weight = proj.weight.data.clone()
+    old_bias = proj.bias.data.clone() if proj.bias is not None else None
+
+    new_proj = nn.Conv2d(
+        in_channels=in_ch,
+        out_channels=proj.out_channels,
+        kernel_size=proj.kernel_size,
+        stride=proj.stride,
+        padding=proj.padding,
+        dilation=proj.dilation,
+        groups=proj.groups,
+        bias=proj.bias is not None,
+        padding_mode=proj.padding_mode,
+    )
+
+    with torch.no_grad():
+        if in_ch <= old_weight.shape[1]:
+            new_proj.weight.copy_(old_weight[:, :in_ch, :, :])
+        else:
+            new_proj.weight[:, :old_weight.shape[1], :, :].copy_(old_weight)
+            extra = in_ch - old_weight.shape[1]
+            mean_w = old_weight.mean(dim=1, keepdim=True)
+            new_proj.weight[:, old_weight.shape[1]:, :, :].copy_(mean_w.repeat(1, extra, 1, 1))
+
+        if old_bias is not None:
+            new_proj.bias.copy_(old_bias)
+
+    segformer_encoder.encoder.patch_embeddings[0].proj = new_proj
+    print(f"[MODEL] patch input channels: 3 -> {in_ch}")
+    return segformer_encoder
+
+
+def replace_segformer_decode_classifier(decode_head: nn.Module, num_classes: int):
+    old = decode_head.classifier
+    if not isinstance(old, nn.Conv2d):
+        raise RuntimeError(f"decode_head.classifier não é Conv2d: {type(old)}")
+
+    new = nn.Conv2d(
+        in_channels=old.in_channels,
+        out_channels=int(num_classes),
+        kernel_size=old.kernel_size,
+        stride=old.stride,
+        padding=old.padding,
+        dilation=old.dilation,
+        groups=old.groups,
+        bias=old.bias is not None,
+        padding_mode=old.padding_mode,
+    )
+    decode_head.classifier = new
+    return decode_head
+
+
+class MultiHeadSegFormer(nn.Module):
+    def __init__(
+        self,
+        backbone: str,
+        channels: int,
+        heads_config: Dict[str, dict],
+        semantic_id2label: Dict[int, str],
+        semantic_label2id: Dict[str, int],
+    ):
+        super().__init__()
+
+        semantic_classes = int(heads_config["semantic"].get("num_classes", len(semantic_id2label)))
+
+        base_config = SegformerConfig.from_pretrained(backbone)
+
+        base_config.num_labels = semantic_classes
+        base_config.id2label = {int(k): str(v) for k, v in semantic_id2label.items()}
+        base_config.label2id = {str(k): int(v) for k, v in semantic_label2id.items()}
+
+        base = SegformerForSemanticSegmentation(base_config)
+
+        patch_segformer_encoder_input_channels(base.segformer, channels)
+        base.config.num_channels = int(channels)
+
+        self.segformer = base.segformer
+        self.decode_heads = nn.ModuleDict()
+        self.heads_config = heads_config
+
+        for head_name, hcfg in heads_config.items():
+            h = copy.deepcopy(base.decode_head)
+            h = replace_segformer_decode_classifier(h, int(hcfg["num_classes"]))
+            self.decode_heads[head_name] = h
+
+        self.config = base.config
+
+    def forward(self, pixel_values: torch.Tensor, head_names=None) -> Dict[str, torch.Tensor]:
+        outputs = self.segformer(
+            pixel_values=pixel_values,
+            output_hidden_states=True,
+            return_dict=True,
+        )
+        hidden_states = outputs.hidden_states
+
+        if head_names is None:
+            selected = list(self.decode_heads.keys())
+        else:
+            selected = [str(h) for h in head_names if str(h) in self.decode_heads]
+
+        return {head_name: self.decode_heads[head_name](hidden_states) for head_name in selected}
+
+
+class MultiSpecSegformerService:
+    """
+    Service de inferência para tensor multiespectral:
+      input : CHW float32 [R,G,B,RE,NIR] 0..1
+      output: dict de predictions multi-head
+    """
+
+    def __init__(self, model_config: dict, mostrar_log=print):
+        self.config = model_config or {}
+        self.mostrar_log = mostrar_log
+
+        self.channels = int(self.config.get("channels", 5))
+        if self.channels != 5:
+            raise RuntimeError(f"MultiSpecSegformerService espera channels=5, veio {self.channels}")
+
+        self.ignore_id = int(self.config.get("ignore_index", 255))
+        self.heads_config = build_heads_config(self.config, ignore_index=self.ignore_id)
+
+        self.semantic_id2label = {
+            0: "chao",
+            1: "cana",
+            2: "erva",
+        }
+
+        self.semantic_label2id = {
+            "chao": 0,
+            "cana": 1,
+            "erva": 2,
+        }
+
+        self.classes = dict(self.semantic_label2id)
+
+        # Dict interno, bom para ids_to_rgb
+        self.colormap_rgb = {
+            0: (85, 85, 85),
+            1: (0, 190, 0),
+            2: (230, 55, 55),
+        }
+
+        # Lista externa, compatível com WeedDetector
+        self.colormap_list_rgb = [
+            self.colormap_rgb[0],
+            self.colormap_rgb[1],
+            self.colormap_rgb[2],
+        ]
+
+        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+        self.use_amp = bool(self.config.get("amp", True)) and self.device.type == "cuda"
+        self.sync_for_timing = bool(self.config.get("sync_for_timing", False)) and self.device.type == "cuda"
+
+        self.runtime_mode = str(self.config.get("runtime_mode", "semantic")).lower()
+        self.lowres_argmax = bool(self.config.get("lowres_argmax", True))
+        self.trust_input = bool(self.config.get("trust_input", True))
+        self.channels_last = bool(self.config.get("channels_last", True)) and self.device.type == "cuda"
+        self.model_half = bool(self.config.get("model_half", False)) and self.device.type == "cuda"
+        
+        if self.device.type == "cuda":
+            torch.backends.cudnn.benchmark = True
+            try:
+                torch.set_float32_matmul_precision("high")
+            except Exception:
+                pass
+
+        self.mean, self.std = self._load_norm_stats_from_config()
+
+        backbone = self.config.get("backbone", self.config.get("pretrained_model", "nvidia/mit-b1"))
+        ckpt_path = self._resolve_checkpoint_path()
+
+        self.mostrar_log(f"[MULTIHEAD] device={self.device}")
+        self.mostrar_log(f"[MULTIHEAD] backbone={backbone}")
+        self.mostrar_log(f"[MULTIHEAD] ckpt={ckpt_path}")
+
+        self.model = MultiHeadSegFormer(
+            backbone=backbone,
+            channels=self.channels,
+            heads_config=self.heads_config,
+            semantic_id2label=self.semantic_id2label,
+            semantic_label2id=self.semantic_label2id,
+        )
+
+        self._load_checkpoint(ckpt_path)
+        self.model.to(self.device)
+        self.model.eval()
+
+        if self.model_half:
+            self.model.half()
+            if self.mean is not None:
+                self.mean = self.mean.half()
+            if self.std is not None:
+                self.std = self.std.half()
+
+        if self.channels_last:
+            try:
+                self.model.to(memory_format=torch.channels_last)
+            except Exception:
+                self.channels_last = False
+
+        if bool(self.config.get("fold_input_norm", False)):
+            self._fold_input_normalization_into_first_conv()
+
+        if bool(self.config.get("torch_compile", False)):
+            try:
+                self.model = torch.compile(
+                    self.model,
+                    mode=str(self.config.get("torch_compile_mode", "reduce-overhead")),
+                    fullgraph=False,
+                )
+                self.mostrar_log("[MULTIHEAD][OPT] torch.compile habilitado")
+            except Exception as e:
+                self.mostrar_log(f"[MULTIHEAD][OPT] torch.compile falhou: {type(e).__name__}: {e}")
+        
+        self.mostrar_log(
+            f"[MULTIHEAD][FAST] runtime_mode={self.runtime_mode} "
+            f"lowres_argmax={self.lowres_argmax} trust_input={self.trust_input} "
+            f"channels_last={self.channels_last} model_half={self.model_half} amp={self.use_amp}"
+        )
+
+        self._ultimo_tensor = None
+        self._ultimo_predictions = None
+        self._ultimo_probs = None
+
+    # ============================================================
+    # Config/load
+    # ============================================================
+
+    def _resolve_checkpoint_path(self) -> Path:
+        candidates = []
+
+        for key in ("ckpt", "checkpoint", "ia_model_path", "model_path"):
+            value = self.config.get(key)
+            if value:
+                candidates.append(Path(value))
+
+        backup_root = self.config.get("backup_root")
+        modelo = self.config.get("modelo", "segformer_b1")
+        model_name = self.config.get("model_name", "test_multi")
+        fusion_mode = self.config.get("fusion_mode", "stacked")
+
+        if backup_root:
+            exp_tag = f"{fusion_mode}_raw{self.channels}_multihead"
+            candidates.extend([
+                Path(backup_root) / modelo / model_name / exp_tag / "best_target.pt",
+                Path(backup_root) / modelo / model_name / exp_tag / "best_score.pt",
+                Path(backup_root) / modelo / model_name / exp_tag / "last.pt",
+            ])
+
+        candidates.extend([
+            Path("backup") / modelo / model_name / f"{fusion_mode}_raw{self.channels}_multihead" / "best_target.pt",
+            Path("backup") / modelo / model_name / f"{fusion_mode}_raw{self.channels}_multihead" / "best_score.pt",
+            Path("backup") / modelo / model_name / f"{fusion_mode}_raw{self.channels}_multihead" / "last.pt",
+        ])
+
+        for p in candidates:
+            p = p.resolve() if not p.is_absolute() else p
+            if p.is_file():
+                return p
+
+        raise FileNotFoundError(
+            "Checkpoint multi-head não encontrado. Procurei:\n" +
+            "\n".join(str(p) for p in candidates)
+        )
+
+    def _load_norm_stats_from_config(self):
+        path = None
+
+        for key in ("norm_stats", "norm_stats_path", "ia_norm_stats_path"):
+            if self.config.get(key):
+                path = Path(self.config[key])
+                break
+
+        if path is None:
+            dataset_root = self.config.get("dataset_root")
+            ia_resolution = self.config.get("ia_resolution", [1024, 640])
+            if dataset_root:
+                w, h = int(ia_resolution[0]), int(ia_resolution[1])
+                path = Path(dataset_root) / f"{w}x{h}" / "group" / "norm_stats.json"
+
+        if path is None or not path.is_file():
+            self.mostrar_log("[MULTIHEAD][NORM] norm_stats não encontrado. Usando tensor 0..1 sem padronização.")
+            return None, None
+
+        js = load_json(path)
+        mean = js.get("mean")
+        std = js.get("std")
+        names = js.get("channels", [])
+
+        if mean is None or std is None:
+            raise RuntimeError(f"norm_stats inválido, faltando mean/std: {path}")
+
+        if len(mean) != self.channels or len(std) != self.channels:
+            raise RuntimeError(
+                f"norm_stats incompatível com channels={self.channels}: "
+                f"mean={len(mean)} std={len(std)} path={path}"
+            )
+
+        self.mostrar_log(f"[MULTIHEAD][NORM] usando {path}")
+        self.mostrar_log(f"[MULTIHEAD][NORM] channels={names}")
+        self.mostrar_log(f"[MULTIHEAD][NORM] mean={mean}")
+        self.mostrar_log(f"[MULTIHEAD][NORM] std ={std}")
+
+        mean_t = torch.tensor(mean, dtype=torch.float32).view(1, self.channels, 1, 1).to(self.device)
+        std_t = torch.tensor(std, dtype=torch.float32).view(1, self.channels, 1, 1).to(self.device)
+
+        return mean_t, std_t
+
+    def _load_checkpoint(self, ckpt_path: Path):
+        ckpt = torch.load(str(ckpt_path), map_location="cpu", weights_only=False)
+
+        if isinstance(ckpt, dict):
+            for key in ("model", "model_state", "model_state_dict", "state_dict"):
+                if key in ckpt and isinstance(ckpt[key], dict):
+                    state = ckpt[key]
+                    break
+            else:
+                state = ckpt
+        else:
+            raise RuntimeError(f"Checkpoint em formato inesperado: {type(ckpt)}")
+
+        clean = {}
+        for k, v in state.items():
+            nk = k
+            for prefix in ("module.", "model."):
+                if nk.startswith(prefix):
+                    nk = nk[len(prefix):]
+            clean[nk] = v
+
+        missing, unexpected = self.model.load_state_dict(clean, strict=False)
+
+        self.mostrar_log(
+            f"[MULTIHEAD] load_state_dict strict=False | "
+            f"missing={len(missing)} unexpected={len(unexpected)}"
+        )
+
+        if missing:
+            self.mostrar_log(f"[MULTIHEAD] primeiros missing: {missing[:8]}")
+        if unexpected:
+            self.mostrar_log(f"[MULTIHEAD] primeiros unexpected: {unexpected[:8]}")
+
+    # ============================================================
+    # Inferência
+    # ============================================================
+
+    def _normalize(self, x: torch.Tensor) -> torch.Tensor:
+        if self.mean is not None and self.std is not None:
+            return (x - self.mean) / torch.clamp(self.std, min=1e-6)
+        return x
+
+    @torch.inference_mode()
+    def infer_tensor(self, tensor5_chw: np.ndarray):
+        """
+        Retorna predictions compatível com o WeedDetector.
+
+        Por enquanto:
+          predictions = semantic mask uint8 HxW
+
+        Também mantém:
+          self._ultimo_predictions_full = dict com semantic/vegetation/cana/target/probs.
+        """
+        if tensor5_chw is None:
+            return None
+
+        chw = np.asarray(tensor5_chw, dtype=np.float32)
+
+        if chw.ndim != 3:
+            raise RuntimeError(f"Tensor inválido: esperado CHW 3D, veio shape={chw.shape}")
+
+        if chw.shape[0] != self.channels:
+            raise RuntimeError(f"Tensor inválido: esperado C={self.channels}, veio shape={chw.shape}")
+
+        chw = np.nan_to_num(chw, nan=0.0, posinf=1.0, neginf=0.0)
+        chw = np.clip(chw, 0.0, 1.0).astype(np.float32, copy=False)
+        chw = np.ascontiguousarray(chw)
+
+        h, w = int(chw.shape[1]), int(chw.shape[2])
+
+        x = torch.from_numpy(chw).unsqueeze(0).to(self.device, non_blocking=True)
+        x = self._normalize(x)
+
+        if self.device.type == "cuda":
+            torch.cuda.synchronize()
+
+        t0 = time.perf_counter()
+
+        with torch.autocast(device_type="cuda", dtype=torch.float16, enabled=self.use_amp):
+            logits_by_head = self.model(pixel_values=x)
+
+            preds = {}
+            probs = {}
+
+            for head_name, logits in logits_by_head.items():
+                logits = F.interpolate(logits, size=(h, w), mode="bilinear", align_corners=False)
+                prob = torch.softmax(logits, dim=1)[0]
+                pred = torch.argmax(prob, dim=0)
+
+                preds[head_name] = pred.detach().cpu().numpy().astype(np.uint8)
+                probs[head_name] = prob.detach().cpu().numpy().astype(np.float32)
+
+        if self.device.type == "cuda":
+            torch.cuda.synchronize()
+
+        t_ms = (time.perf_counter() - t0) * 1000.0
+
+        semantic = preds["semantic"]
+        vegetation = preds["vegetation"]
+        cana = preds["cana"]
+
+        target = self.operational_target_mask(vegetation, cana, ignore_id=self.ignore_id)
+
+        self._ultimo_tensor = chw
+        self._ultimo_predictions = semantic
+        self._ultimo_probs = probs
+        self._ultimo_predictions_full = {
+            "semantic": semantic,
+            "vegetation": vegetation,
+            "cana": cana,
+            "target": target,
+            "probs": probs,
+            "infer_ms": t_ms,
+        }
+
+        return semantic
+
+    def infer_tensor_full(self, tensor5_chw: np.ndarray):
+        semantic = self.infer_tensor(tensor5_chw)
+        if semantic is None:
+            return None
+        return dict(self._ultimo_predictions_full)
+
+    @torch.inference_mode()
+    def infer_tensor_fast(self, tensor5_chw: np.ndarray, keep_probs: bool = False):
+        if keep_probs:
+            return self.infer_tensor(tensor5_chw)
+
+        return self.infer_tensor_ultrafast(
+            tensor5_chw,
+            return_full=bool(self.config.get("return_full_fast", False)),
+        )
+
+    # ============================================================
+    # Preview/debug
+    # ============================================================
+
+    def preview_infer_cached(self, tensor5_chw=None, predictions=None, alpha=0.5):
+        """
+        Compatível com o uso atual do weed_worker:
+          rgb_frame, seg_frame, overlay_frame, _, _ = preview_infer_cached(...)
+
+        Retorna BGR para OpenCV/TCP.
+        """
+        tensor = tensor5_chw if tensor5_chw is not None else self._ultimo_tensor
+        pred = predictions if predictions is not None else self._ultimo_predictions
+
+        if tensor is None:
+            return None, None, None, None, None
+
+        rgb = self.tensor_to_preview_rgb(tensor)
+        rgb_bgr = cv2.cvtColor(rgb, cv2.COLOR_RGB2BGR)
+
+        if pred is None:
+            return rgb_bgr, None, rgb_bgr, None, None
+
+        seg_rgb = self.ids_to_rgb(pred, self.colormap_rgb, ignore_id=self.ignore_id)
+        overlay_rgb = cv2.addWeighted(rgb, 1.0 - alpha, seg_rgb, alpha, 0.0)
+
+        seg_bgr = cv2.cvtColor(seg_rgb, cv2.COLOR_RGB2BGR)
+        overlay_bgr = cv2.cvtColor(overlay_rgb, cv2.COLOR_RGB2BGR)
+
+        return rgb_bgr, seg_bgr, overlay_bgr, None, None
+
+    def get_classes(self):
+        return self.classes
+
+    def get_colormap(self):
+        return self.colormap_list_rgb
+
+    @staticmethod
+    def operational_target_mask(veg_mask: np.ndarray, cana_mask: np.ndarray, ignore_id: int = 255) -> np.ndarray:
+        out = np.zeros_like(veg_mask, dtype=np.uint8)
+        ignore = (veg_mask == ignore_id) | (cana_mask == ignore_id)
+        out[(veg_mask == 1) & (cana_mask == 0)] = 1
+        out[ignore] = ignore_id
+        return out
+
+    @staticmethod
+    def tensor_to_preview_rgb(chw: np.ndarray, gamma: float = 0.85) -> np.ndarray:
+        c, h, w = chw.shape
+
+        if c >= 3:
+            rgb = np.transpose(chw[:3], (1, 2, 0)).copy()
+        else:
+            one = chw[0]
+            rgb = np.stack([one, one, one], axis=-1)
+
+        rgb = np.nan_to_num(rgb, nan=0.0, posinf=1.0, neginf=0.0)
+
+        lo = np.percentile(rgb, 1.0)
+        hi = np.percentile(rgb, 99.0)
+
+        if hi > lo:
+            rgb = (rgb - lo) / (hi - lo)
+
+        rgb = np.clip(rgb, 0.0, 1.0)
+
+        if gamma and gamma > 0:
+            rgb = np.power(rgb, gamma)
+
+        return (rgb * 255.0).astype(np.uint8)
+
+    @staticmethod
+    def ids_to_rgb(mask: np.ndarray, colormap_rgb: Dict[int, Tuple[int, int, int]], ignore_id: int = 255) -> np.ndarray:
+        h, w = mask.shape[:2]
+        out = np.zeros((h, w, 3), dtype=np.uint8)
+
+        for cid, color in colormap_rgb.items():
+            out[mask == cid] = color
+
+        out[mask == ignore_id] = (0, 0, 0)
+        return out
+    
+
+
+    def _runtime_head_names(self):
+        mode = str(getattr(self, "runtime_mode", "semantic")).lower()
+
+        if mode in ("semantic", "sem", "mask"):
+            return ["semantic"]
+
+        if mode in ("target", "spray", "operational"):
+            return ["vegetation", "cana"]
+
+        if mode in ("target_direct", "direct_target", "target_head"):
+            return ["target"]
+
+        if mode in ("all", "full", "debug"):
+            return ["semantic", "vegetation", "cana", "target"]
+
+        return ["semantic"]
+
+    def _prepare_input_tensor_fast(self, tensor5_chw: np.ndarray):
+        if tensor5_chw is None:
+            return None, None, None
+
+        # O pipeline científico já devolve float32 CHW 0..1 contíguo.
+        # Em runtime, confiar nisso economiza uma passada grande no CPU.
+        if bool(getattr(self, "trust_input", True)):
+            chw = tensor5_chw
+            if not isinstance(chw, np.ndarray):
+                chw = np.asarray(chw, dtype=np.float32)
+            if chw.dtype != np.float32 or not chw.flags.c_contiguous:
+                chw = np.ascontiguousarray(chw, dtype=np.float32)
+        else:
+            chw = np.asarray(tensor5_chw, dtype=np.float32)
+            chw = np.nan_to_num(chw, nan=0.0, posinf=1.0, neginf=0.0)
+            chw = np.clip(chw, 0.0, 1.0).astype(np.float32, copy=False)
+            chw = np.ascontiguousarray(chw)
+
+        if chw.ndim != 3:
+            raise RuntimeError(f"Tensor inválido: esperado CHW 3D, veio shape={chw.shape}")
+
+        if chw.shape[0] != self.channels:
+            raise RuntimeError(f"Tensor inválido: esperado C={self.channels}, veio shape={chw.shape}")
+
+        h, w = int(chw.shape[1]), int(chw.shape[2])
+
+        x = torch.from_numpy(chw).unsqueeze(0).to(self.device, non_blocking=True)
+
+        if bool(getattr(self, "channels_last", False)):
+            try:
+                x = x.contiguous(memory_format=torch.channels_last)
+            except Exception:
+                pass
+
+        x = self._normalize(x)
+
+        if bool(getattr(self, "model_half", False)):
+            x = x.half()
+
+        return chw, x, (h, w)
+
+    def _logits_to_pred_numpy_fast(self, logits: torch.Tensor, out_hw, lowres_argmax: bool):
+        h, w = int(out_hw[0]), int(out_hw[1])
+
+        if lowres_argmax:
+            # Argmax no mapa pequeno, depois resize da máscara uint8.
+            # Bem mais barato que interpolar logits CxHxW em float.
+            pred_small = torch.argmax(logits, dim=1)[0]
+            pred_np = pred_small.detach().to("cpu", non_blocking=False).numpy().astype(np.uint8)
+
+            if pred_np.shape[0] != h or pred_np.shape[1] != w:
+                pred_np = cv2.resize(pred_np, (w, h), interpolation=cv2.INTER_NEAREST)
+
+            return pred_np
+
+        # Caminho equivalente ao atual, mas sem softmax.
+        logits = F.interpolate(logits, size=(h, w), mode="bilinear", align_corners=False)
+        pred = torch.argmax(logits, dim=1)[0]
+        return pred.detach().to("cpu", non_blocking=False).numpy().astype(np.uint8)
+
+    @torch.inference_mode()
+    def infer_tensor_ultrafast(self, tensor5_chw: np.ndarray, return_full: bool = False):
+        if tensor5_chw is None:
+            return None
+
+        t_total0 = time.perf_counter()
+
+        # ------------------------------------------------------------
+        # Prepare: numpy -> torch/cuda + layout + normalização
+        # ------------------------------------------------------------
+        t_prepare0 = time.perf_counter()
+
+        chw, x, out_hw = self._prepare_input_tensor_fast(tensor5_chw)
+
+        prepare_ms = (time.perf_counter() - t_prepare0) * 1000.0
+
+        if x is None:
+            return None
+
+        head_names = self._runtime_head_names()
+        lowres_argmax = bool(getattr(self, "lowres_argmax", True))
+        runtime_mode = str(getattr(self, "runtime_mode", "semantic")).lower()
+        output_mask_fullres = bool(self.config.get("output_mask_fullres", True))
+
+        if bool(getattr(self, "sync_for_timing", False)) and self.device.type == "cuda":
+            torch.cuda.synchronize()
+
+        # ------------------------------------------------------------
+        # Forward: modelo UMA vez só
+        # ------------------------------------------------------------
+        t_forward0 = time.perf_counter()
+
+        with torch.autocast(
+            device_type="cuda",
+            dtype=torch.float16,
+            enabled=bool(getattr(self, "use_amp", True)) and self.device.type == "cuda",
+        ):
+            logits_by_head = self.model(pixel_values=x, head_names=head_names)
+
+        if self.device.type == "cuda":
+            torch.cuda.synchronize()
+
+        forward_ms = (time.perf_counter() - t_forward0) * 1000.0
+
+        # ------------------------------------------------------------
+        # Post: argmax / target / cópia CPU
+        # ------------------------------------------------------------
+        t_post0 = time.perf_counter()
+
+        semantic = None
+        vegetation = None
+        cana = None
+        target = None
+
+        if runtime_mode in ("target", "spray", "operational"):
+            logits_veg = logits_by_head.get("vegetation")
+            logits_cana = logits_by_head.get("cana")
+
+            if logits_veg is None or logits_cana is None:
+                raise RuntimeError("runtime_mode=target requer heads vegetation e cana")
+
+            target = self._target_from_logits_gpu_fast(
+                logits_veg,
+                logits_cana,
+                out_hw=out_hw,
+                fullres=output_mask_fullres,
+            )
+
+            output = target
+
+        else:
+            preds = {}
+
+            if "target" in logits_by_head:
+                target = self._logits_to_pred_numpy_fast(
+                    logits_by_head["target"],
+                    out_hw=out_hw,
+                    lowres_argmax=lowres_argmax,
+                )
+
+                semantic = None
+                vegetation = None
+                cana = None
+                output = target
+            else:
+                for head_name, logits in logits_by_head.items():
+                    preds[head_name] = self._logits_to_pred_numpy_fast(
+                        logits,
+                        out_hw=out_hw,
+                        lowres_argmax=lowres_argmax,
+                    )
+
+                semantic = preds.get("semantic")
+                vegetation = preds.get("vegetation")
+                cana = preds.get("cana")
+
+                if vegetation is not None and cana is not None:
+                    target = self.operational_target_mask(
+                        vegetation,
+                        cana,
+                        ignore_id=self.ignore_id,
+                    )
+
+                output = semantic if semantic is not None else target
+
+        post_ms = (time.perf_counter() - t_post0) * 1000.0
+        total_ms = (time.perf_counter() - t_total0) * 1000.0
+
+        self._ultimo_tensor = chw
+        self._ultimo_predictions = output
+        self._ultimo_probs = None
+        self._ultimo_predictions_full = {
+            "semantic": semantic,
+            "vegetation": vegetation,
+            "cana": cana,
+            "target": target,
+            "probs": None,
+            "infer_ms": total_ms,
+            "prepare_ms": prepare_ms,
+            "forward_ms": forward_ms,
+            "post_ms": post_ms,
+            "runtime_mode": runtime_mode,
+            "lowres_argmax": lowres_argmax,
+            "heads": list(head_names),
+            "output_mask_fullres": output_mask_fullres,
+        }
+
+        if return_full:
+            return dict(self._ultimo_predictions_full)
+
+        return output
+
+
+    def _target_from_logits_gpu_fast(self, logits_veg, logits_cana, out_hw, fullres: bool = True):
+        h, w = int(out_hw[0]), int(out_hw[1])
+
+        veg = torch.argmax(logits_veg, dim=1)[0]
+        cana = torch.argmax(logits_cana, dim=1)[0]
+
+        target = ((veg == 1) & (cana == 0)).to(torch.uint8)
+
+        target_np = target.detach().to("cpu", non_blocking=False).numpy()
+
+        if fullres and (target_np.shape[0] != h or target_np.shape[1] != w):
+            target_np = cv2.resize(
+                target_np,
+                (w, h),
+                interpolation=cv2.INTER_NEAREST,
+            )
+
+        return target_np.astype(np.uint8, copy=False)
+
+    def _fold_input_normalization_into_first_conv(self):
+        if self.mean is None or self.std is None:
+            return False
+
+        try:
+            proj = self.model.segformer.encoder.patch_embeddings[0].proj
+        except Exception:
+            return False
+
+        if not isinstance(proj, nn.Conv2d):
+            return False
+
+        with torch.no_grad():
+            device = proj.weight.device
+            dtype = proj.weight.dtype
+
+            mean = self.mean.detach().to(device=device, dtype=dtype).view(-1)
+            std = self.std.detach().to(device=device, dtype=dtype).view(-1)
+            std = torch.clamp(std, min=1e-6)
+
+            w_old = proj.weight.data.clone()
+            b_old = proj.bias.data.clone() if proj.bias is not None else torch.zeros(
+                proj.out_channels,
+                device=device,
+                dtype=dtype,
+            )
+
+            # W'[:, c] = W[:, c] / std[c]
+            w_new = w_old / std.view(1, -1, 1, 1)
+
+            # bias' = bias - sum(W[:,c,:,:] * mean[c] / std[c])
+            offset = (w_old * (mean / std).view(1, -1, 1, 1)).sum(dim=(1, 2, 3))
+            b_new = b_old - offset
+
+            proj.weight.data.copy_(w_new)
+
+            if proj.bias is None:
+                proj.bias = nn.Parameter(b_new)
+            else:
+                proj.bias.data.copy_(b_new)
+
+        self.mean = None
+        self.std = None
+
+        self.mostrar_log("[MULTIHEAD][OPT] normalização foldada na primeira conv")
+        return True
+

Python/OAK/datasets/oak-fcc-3/core/test.json

@@ -0,0 +1,99 @@
+{
+            "debug_visual": false,
+            "frames_consecutivos": 3,
+            "frames_histerese": 2,
+            "min_area_px": 400, 
+            "max_area_frac": 0.2,
+            "ia_roi_begin": 0.0,
+            "ia_roi_size": 1.0,
+            "ia_resolution": [1024,640],
+            "ia_channels": 5,
+            "ia_use_ndvi": false,
+            "erva_top_band_frac": 0.30,
+            "erva_frac_ema": 0.3,
+            "erva_thresh_vel_gain": 0.4,
+            "min_frac_erva_global_on": 0.0020,
+            "min_frac_erva_global_off": 0.0015,
+            "min_frac_erva_top_on": 0.0015,
+            "min_frac_erva_top_off": 0.0010,
+            "min_frac_erva_por_bico": 0.02,
+            "usar_morfologia": true,
+            "kernel_morf": 3,
+
+            "usar_radar_global_gate": true,
+            "max_frac_cana_por_bico": 0.009,
+            "ema_frac_bico": 0.35,
+            "on_frames_required": 3,
+            "off_frames_required": 2,
+            "cana_halo_px": 5,
+            "min_area_erva_px": 80,
+            "erva_thresh_vel_gain_local": 0.6,
+            "k_roi_shift_px_per_vnorm": 24.0,
+
+            "fps": 20,
+            "module_params": "C:\\ZendionInc\\agrobot_base\\Python\\OAK\\datasets\\oak-fcc-3\\calibration\\module_params.json",
+
+            "qtd_bicos": 7,
+            "velocidade_robo": 0,
+            "ia_model_path": "C:\\AgroBaseModels\\Ervas\\model-3_1.pt",
+            "ia_labelmap_path": "C:\\AgroBaseModels\\Ervas\\model-3_1.txt",
+            "ia_norm_stats_path": "C:\\AgroBaseModels\\Ervas\\model-1_1.json",
+            "ia_backbone": "nvidia/mit-b1",
+
+            "faixa_atuacao_bicos": 0.7,
+            "area_atuacao_bicos": 0.1,
+            "min_frac_erva_por_bico_on": 0.02,
+            "min_frac_erva_por_bico_off": 0.01,
+
+            "tipo_camera_solo": "multispectral",
+            "channels": 5,
+            "backbone": "nvidia/mit-b1",
+            "ckpt": "C:\\AgroBaseModels\\Ervas\\model-3_1.pt",
+            "norm_stats_path": "C:\\AgroBaseModels\\Ervas\\model-1_1.json",
+            "module_calibration_json": "C:\\ZendionInc\\agrobot_base\\Python\\OAK\\datasets\\oak-fcc-3\\calibration\\module_params.json",
+            "camera_width": 1280,
+            "camera_height": 800,
+            "camera_fps": 40,
+            "amp": true,
+            "fold_input_norm": true,
+            "runtime_mode": "target_direct",
+            "output_mask_fullres": false,
+            "lowres_argmax": true,
+            "trust_input": true,
+            "channels_last": false,
+            "model_half": true,
+            "sync_for_timing": false,
+            "torch_compile": false,
+            "torch_compile_mode": "reduce-overhead",
+            "heads": {
+                "semantic": {
+                    "enabled": true,
+                    "type": "multiclass",
+                    "num_classes": 3,
+                    "classes": {"chao": 0, "cana": 1, "erva": 2},
+                    "ignore_index": 255
+                },
+                "vegetation": {
+                    "enabled": true,
+                    "type": "binary",
+                    "num_classes": 2,
+                    "classes": {"background": 0, "vegetation": 1},
+                    "ignore_index": 255
+                },
+                "cana": {
+                    "enabled": true,
+                    "type": "binary",
+                    "num_classes": 2,
+                    "classes": {"not_cana": 0, "cana": 1},
+                    "ignore_index": 255
+                },
+                "target": {
+                    "enabled": true,
+                    "type": "binary",
+                    "num_classes": 2,
+                    "classes": {"background": 0, "target": 1},
+                    "ignore_index": 255
+                }
+            }
+
+        }

Python/OAK/datasets/oak-fcc-3/dataset/labelmap.txt

@@ -0,0 +1,5 @@
+# label:color_rgb:parts:actions
+chao:128,0,0::
+cana:0,0,128::
+erva:0,128,0::
+ignore:255,255,255::

Python/OAK/datasets/oak-fcc-3/tests/test.py

@@ -1,37 +1,68 @@
-from core.oak_fcc3_service import OakFcc3Service
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
 
-svc = OakFcc3Service(
-    fps=15,
-    width=640,
-    height=400,
-    frame_type="MULTISPEC",
-    capture_mode="TRIPLE",
-    raw_policy="require_triple",
-    sync_mode="best",
-    sync_tolerance_ms=25.0,
-)
+import depthai as dai
 
-svc.connect()
-svc.begin(frame_type="RAW_BRUTO", output_dtype="uint8", capture_mode="TRIPLE")
 
-frame, meta = svc.capture_frame(timeout=2.0)
+def ler_attr_ou_metodo(obj, nomes, default=None):
+    for nome in nomes:
+        try:
+            valor = getattr(obj, nome, None)
 
-print(meta["camera_info"])
-for cam_id, arr in frame.items():
-    print(cam_id, arr.shape, arr.dtype, arr.size)
+            if valor is None:
+                continue
 
-from core.raw_processor_core import RawProcessorCore
-from core.raw_processor_preview import RawProcessorPreview
-import cv2
+            if callable(valor):
+                return valor()
 
-core = RawProcessorCore(sensor_width=1280, sensor_height=800, bayer_pattern="GBRG")
-preview = RawProcessorPreview(sensor_width=1280, sensor_height=800, bayer_pattern="GBRG")
+            return valor
+        except Exception:
+            pass
 
-raw16 = core.unpack_raw10_packed(frame["cam0"], sensor_width=1280, sensor_height=800)
-img = preview.raw16_to_preview_bgr(raw16, bit_depth=10)
+    return default
 
-cv2.imwrite("calibration/cam0_raw_preview.png", img)
-print(raw16.shape, raw16.dtype, raw16.min(), raw16.max())
 
-svc.stop()
-svc.disconnect()
+def main():
+    print("==========================================")
+    print("Teste simples - listar dispositivos DepthAI")
+    print("==========================================")
+
+    try:
+        devices = dai.Device.getAllAvailableDevices()
+    except Exception as e:
+        print("[ERRO] Falha ao chamar dai.Device.getAllAvailableDevices()")
+        print("Erro:", e)
+        return
+
+    if not devices:
+        print("[INFO] Nenhum dispositivo DepthAI/OAK encontrado.")
+        return
+
+    print(f"[OK] Dispositivos encontrados: {len(devices)}")
+    print()
+
+    for i, dev_info in enumerate(devices, start=1):
+        device_id = ler_attr_ou_metodo(
+            dev_info,
+            ["getMxId", "mxid", "getDeviceId", "deviceId"],
+            default=None
+        )
+
+        name = ler_attr_ou_metodo(dev_info, ["name"], default=None)
+        state = ler_attr_ou_metodo(dev_info, ["state"], default=None)
+        protocol = ler_attr_ou_metodo(dev_info, ["protocol"], default=None)
+        platform = ler_attr_ou_metodo(dev_info, ["platform"], default=None)
+        status = ler_attr_ou_metodo(dev_info, ["status"], default=None)
+
+        print(f"Dispositivo #{i}")
+        print(f"  DeviceId : {device_id}")
+        print(f"  Name     : {name}")
+        print(f"  State    : {state}")
+        print(f"  Protocol : {protocol}")
+        print(f"  Platform : {platform}")
+        print(f"  Status   : {status}")
+        print("-" * 42)
+
+
+if __name__ == "__main__":
+    main()

Python/OAK/datasets/oak-fcc-3/tests/test_oak_fcc3_aligned_geometry.py

@@ -0,0 +1,534 @@
+import argparse
+import json
+import time
+from pathlib import Path
+
+import cv2
+import depthai as dai
+import numpy as np
+
+
+# ============================================================
+# Homografia / crop helpers
+# ============================================================
+
+def load_module_params(path: str) -> dict:
+    with open(path, "r", encoding="utf-8") as f:
+        return json.load(f)
+
+
+def scale_homography_to_runtime(H, calib_size, runtime_size):
+    """
+    Mesmo conceito do RawProcessorCore:
+      H calib: ponto_spec_calib -> ponto_rgb_calib
+      H runtime: ponto_spec_runtime -> ponto_rgb_runtime
+    """
+    H = np.asarray(H, dtype=np.float32)
+
+    if calib_size is None:
+        if abs(H[2, 2]) > 1e-9:
+            H = H / H[2, 2]
+        return H.astype(np.float32)
+
+    calib_w, calib_h = calib_size
+    runtime_w, runtime_h = runtime_size
+
+    calib_w = float(calib_w)
+    calib_h = float(calib_h)
+    runtime_w = float(runtime_w)
+    runtime_h = float(runtime_h)
+
+    sx = runtime_w / calib_w
+    sy = runtime_h / calib_h
+
+    S = np.array(
+        [
+            [sx, 0.0, 0.0],
+            [0.0, sy, 0.0],
+            [0.0, 0.0, 1.0],
+        ],
+        dtype=np.float32,
+    )
+
+    S_inv = np.array(
+        [
+            [1.0 / sx, 0.0, 0.0],
+            [0.0, 1.0 / sy, 0.0],
+            [0.0, 0.0, 1.0],
+        ],
+        dtype=np.float32,
+    )
+
+    H_runtime = S @ H @ S_inv
+
+    if abs(H_runtime[2, 2]) > 1e-9:
+        H_runtime = H_runtime / H_runtime[2, 2]
+
+    return H_runtime.astype(np.float32)
+
+
+def warp_mask(mask, H, out_w, out_h):
+    return cv2.warpPerspective(
+        mask,
+        H,
+        (out_w, out_h),
+        flags=cv2.INTER_NEAREST,
+        borderMode=cv2.BORDER_CONSTANT,
+        borderValue=0,
+    )
+
+
+def compute_common_crop_box(runtime_w, runtime_h, H_re, H_nir):
+    """
+    Calcula no PC uma vez só a área comum válida no espaço RGB.
+    Depois essa área vira o retângulo final usado na OAK.
+    """
+    base = np.ones((runtime_h, runtime_w), dtype=np.uint8) * 255
+
+    rgb_mask = base
+    re_mask = warp_mask(base, H_re, runtime_w, runtime_h)
+    nir_mask = warp_mask(base, H_nir, runtime_w, runtime_h)
+
+    common = (rgb_mask > 0) & (re_mask > 0) & (nir_mask > 0)
+
+    ys, xs = np.where(common)
+    if xs.size == 0 or ys.size == 0:
+        raise RuntimeError("Área comum vazia. Verifique as homografias.")
+
+    x0 = int(xs.min())
+    x1 = int(xs.max()) + 1
+    y0 = int(ys.min())
+    y1 = int(ys.max()) + 1
+
+    return x0, y0, x1, y1
+
+
+def apply_H_to_point(H, x, y):
+    p = np.array([float(x), float(y), 1.0], dtype=np.float32)
+    q = H @ p
+    if abs(q[2]) < 1e-9:
+        return float(q[0]), float(q[1])
+    return float(q[0] / q[2]), float(q[1] / q[2])
+
+
+def quad_for_output_crop_to_input(H_src_to_rgb, crop_box):
+    """
+    A OAK/ImageManip recebe uma quadrilateral SOURCE que será esticada
+    para o retângulo de saída.
+
+    Como nossa homografia original é source -> RGB, usamos H_inv para descobrir:
+      canto final no espaço RGB -> ponto correspondente no source.
+    """
+    x0, y0, x1, y1 = crop_box
+
+    dst_corners_rgb = [
+        (x0, y0),  # top-left
+        (x1, y0),  # top-right
+        (x1, y1),  # bottom-right
+        (x0, y1),  # bottom-left
+    ]
+
+    H_inv = np.linalg.inv(H_src_to_rgb).astype(np.float32)
+
+    src_quad = []
+    for x, y in dst_corners_rgb:
+        sx, sy = apply_H_to_point(H_inv, x, y)
+        src_quad.append((sx, sy))
+
+    return src_quad
+
+
+def identity_quad_for_crop(crop_box):
+    x0, y0, x1, y1 = crop_box
+    return [
+        (float(x0), float(y0)),
+        (float(x1), float(y0)),
+        (float(x1), float(y1)),
+        (float(x0), float(y1)),
+    ]
+
+
+def dai_points(quad):
+    return [dai.Point2f(float(x), float(y)) for x, y in quad]
+
+
+# ============================================================
+# DepthAI pipeline helpers
+# ============================================================
+
+def create_color_camera(pipeline, socket, width, height, fps):
+    cam = pipeline.create(dai.node.ColorCamera)
+    cam.setBoardSocket(socket)
+
+    # OV9782 costuma ser 800p. Se sua versão não aceitar, ajuste para a resolução suportada.
+    cam.setResolution(dai.ColorCameraProperties.SensorResolution.THE_800_P)
+
+    cam.setFps(float(fps))
+    cam.setInterleaved(False)
+    cam.setColorOrder(dai.ColorCameraProperties.ColorOrder.BGR)
+    cam.setVideoSize(int(width), int(height))
+
+    return cam
+
+
+def create_mono_camera(pipeline, socket, width, height, fps):
+    cam = pipeline.create(dai.node.MonoCamera)
+    cam.setBoardSocket(socket)
+
+    # OV9282 normalmente suporta 800p.
+    cam.setResolution(dai.MonoCameraProperties.SensorResolution.THE_800_P)
+    cam.setFps(float(fps))
+
+    return cam
+
+
+def create_warp_manip(
+    pipeline,
+    name,
+    out_w,
+    out_h,
+    src_quad_px,
+    frame_type=None,
+    max_output_frame_size=None,
+):
+    manip = pipeline.create(dai.node.ImageManip)
+
+    dst_quad_px = [
+        (0.0, 0.0),
+        (float(out_w - 1), 0.0),
+        (float(out_w - 1), float(out_h - 1)),
+        (0.0, float(out_h - 1)),
+    ]
+
+    # Faz crop/warp/resize em uma operação:
+    # src_quad_px na imagem original -> dst_quad_px na imagem final.
+    manip.initialConfig.addTransformFourPoints(
+        dai_points(src_quad_px),
+        dai_points(dst_quad_px),
+        False,
+    )
+
+    # NÃO usar setResize nessa versão do DepthAI.
+    # A saída já deve sair com o tamanho definido pelo dst_quad_px.
+    manip.setMaxOutputFrameSize(
+        int(max_output_frame_size)
+        if max_output_frame_size is not None
+        else int(out_w * out_h * 3)
+    )
+
+    if frame_type is not None:
+        manip.initialConfig.setFrameType(frame_type)
+
+    xout = pipeline.createXLinkOut()
+    xout.setStreamName(name)
+
+    manip.out.link(xout.input)
+
+    return manip, xout
+
+
+def drain_latest(queue):
+    """
+    Pega o frame mais novo disponível sem acumular fila.
+    """
+    latest = None
+
+    while True:
+        msg = queue.tryGet()
+        if msg is None:
+            break
+        latest = msg
+
+    return latest
+
+
+def clamp_quad(quad, w, h):
+    out = []
+    for x, y in quad:
+        x = max(0.0, min(float(w - 1), float(x)))
+        y = max(0.0, min(float(h - 1), float(y)))
+        out.append((x, y))
+    return out
+
+
+# ============================================================
+# Main benchmark
+# ============================================================
+
+def main():
+    ap = argparse.ArgumentParser(
+        description="Teste OAK-FFC-3: resize/crop/homografia na OAK e envio de frames alinhados."
+    )
+
+    ap.add_argument("--mxid", default=None)
+    ap.add_argument("--module_params", required=True)
+    ap.add_argument("--fps", type=float, default=20.0)
+
+    ap.add_argument("--sensor_w", type=int, default=1280)
+    ap.add_argument("--sensor_h", type=int, default=800)
+
+    ap.add_argument("--out_w", type=int, default=1024)
+    ap.add_argument("--out_h", type=int, default=640)
+
+    ap.add_argument("--seconds", type=float, default=20.0)
+    ap.add_argument("--save_debug", action="store_true")
+    ap.add_argument("--debug_dir", default="oak_aligned_debug")
+
+    args = ap.parse_args()
+
+    params = load_module_params(args.module_params)
+    fusion = params.get("fusion_config", {}) or {}
+
+    homographies = fusion.get("homographies", {}) or {}
+    H_re_raw = homographies.get("re_to_rgb")
+    H_nir_raw = homographies.get("nir_to_rgb")
+
+    if H_re_raw is None or H_nir_raw is None:
+        raise RuntimeError("module_params precisa conter fusion_config.homographies.re_to_rgb e nir_to_rgb")
+
+    calib_size = fusion.get("homography_calibration_size", None)
+    runtime_size = (args.sensor_w, args.sensor_h)
+
+    H_re = scale_homography_to_runtime(H_re_raw, calib_size, runtime_size)
+    H_nir = scale_homography_to_runtime(H_nir_raw, calib_size, runtime_size)
+
+    crop_box = compute_common_crop_box(args.sensor_w, args.sensor_h, H_re, H_nir)
+
+    quad_rgb = identity_quad_for_crop(crop_box)
+    quad_re = quad_for_output_crop_to_input(H_re, crop_box)
+    quad_nir = quad_for_output_crop_to_input(H_nir, crop_box)
+
+    quad_rgb = clamp_quad(quad_rgb, args.sensor_w, args.sensor_h)
+    quad_re = clamp_quad(quad_re, args.sensor_w, args.sensor_h)
+    quad_nir = clamp_quad(quad_nir, args.sensor_w, args.sensor_h)
+
+    print("============================================")
+    print("OAK-FCC-3 aligned geometry benchmark")
+    print(f"module_params : {args.module_params}")
+    print(f"sensor        : {args.sensor_w}x{args.sensor_h}")
+    print(f"output        : {args.out_w}x{args.out_h}")
+    print(f"fps target    : {args.fps}")
+    print(f"crop_box RGB  : {crop_box}")
+    print(f"quad_rgb      : {quad_rgb}")
+    print(f"quad_re       : {quad_re}")
+    print(f"quad_nir      : {quad_nir}")
+    print("============================================")
+
+    pipeline = dai.Pipeline()
+
+    # Sockets assumidos:
+    # CAM_A = RGB, CAM_B = RE, CAM_C = NIR
+    rgb = create_color_camera(
+        pipeline,
+        dai.CameraBoardSocket.CAM_A,
+        args.sensor_w,
+        args.sensor_h,
+        args.fps,
+    )
+
+    re = create_mono_camera(
+        pipeline,
+        dai.CameraBoardSocket.CAM_B,
+        args.sensor_w,
+        args.sensor_h,
+        args.fps,
+    )
+
+    nir = create_mono_camera(
+        pipeline,
+        dai.CameraBoardSocket.CAM_C,
+        args.sensor_w,
+        args.sensor_h,
+        args.fps,
+    )
+
+    # Tamanhos aproximados de saída.
+    # RGB BGR888p = 3 canais. Mono GRAY8 = 1 canal.
+    rgb_manip, _ = create_warp_manip(
+        pipeline,
+        "rgb_aligned",
+        args.out_w,
+        args.out_h,
+        quad_rgb,
+        frame_type=dai.ImgFrame.Type.BGR888p,
+        max_output_frame_size=args.out_w * args.out_h * 3,
+    )
+
+    re_manip, _ = create_warp_manip(
+        pipeline,
+        "re_aligned",
+        args.out_w,
+        args.out_h,
+        quad_re,
+        frame_type=dai.ImgFrame.Type.GRAY8,
+        max_output_frame_size=args.out_w * args.out_h,
+    )
+
+    nir_manip, _ = create_warp_manip(
+        pipeline,
+        "nir_aligned",
+        args.out_w,
+        args.out_h,
+        quad_nir,
+        frame_type=dai.ImgFrame.Type.GRAY8,
+        max_output_frame_size=args.out_w * args.out_h,
+    )
+
+    # Links.
+    # RGB usa video, RE/NIR usam mono out.
+    rgb.video.link(rgb_manip.inputImage)
+    re.out.link(re_manip.inputImage)
+    nir.out.link(nir_manip.inputImage)
+
+    device_info = None
+    if args.mxid:
+        device_info = dai.DeviceInfo(args.mxid)
+
+    if args.save_debug:
+        Path(args.debug_dir).mkdir(parents=True, exist_ok=True)
+
+    with dai.Device(pipeline, device_info) if device_info is not None else dai.Device(pipeline) as device:
+        q_rgb = device.getOutputQueue("rgb_aligned", maxSize=2, blocking=False)
+        q_re = device.getOutputQueue("re_aligned", maxSize=2, blocking=False)
+        q_nir = device.getOutputQueue("nir_aligned", maxSize=2, blocking=False)
+
+        t_start = time.perf_counter()
+        t_last_log = t_start
+
+        n_triplets = 0
+        n_rgb = 0
+        n_re = 0
+        n_nir = 0
+
+        last_rgb = None
+        last_re = None
+        last_nir = None
+
+        lat_samples_ms = []
+        stack_samples_ms = []
+
+        saved = False
+
+        while True:
+            now = time.perf_counter()
+            if now - t_start >= args.seconds:
+                break
+
+            msg_rgb = drain_latest(q_rgb)
+            msg_re = drain_latest(q_re)
+            msg_nir = drain_latest(q_nir)
+
+            if msg_rgb is not None:
+                last_rgb = msg_rgb
+                n_rgb += 1
+
+            if msg_re is not None:
+                last_re = msg_re
+                n_re += 1
+
+            if msg_nir is not None:
+                last_nir = msg_nir
+                n_nir += 1
+
+            if last_rgb is not None and last_re is not None and last_nir is not None:
+                # Mede latência aproximada usando timestamps do dispositivo.
+                try:
+                    ts_dev = [
+                        last_rgb.getTimestampDevice().total_seconds(),
+                        last_re.getTimestampDevice().total_seconds(),
+                        last_nir.getTimestampDevice().total_seconds(),
+                    ]
+                    sync_dt_ms = (max(ts_dev) - min(ts_dev)) * 1000.0
+                    lat_samples_ms.append(sync_dt_ms)
+                except Exception:
+                    sync_dt_ms = -1.0
+
+                # Mede custo de converter e empilhar no PC.
+                t_stack0 = time.perf_counter()
+
+                rgb_bgr = last_rgb.getCvFrame()
+                re_gray = last_re.getCvFrame()
+                nir_gray = last_nir.getCvFrame()
+
+                # Só para simular o stack final.
+                # RGB vem BGR, então convertemos para RGB.
+                rgb_float = cv2.cvtColor(rgb_bgr, cv2.COLOR_BGR2RGB).astype(np.float32) / 255.0
+                re_float = re_gray.astype(np.float32) / 255.0
+                nir_float = nir_gray.astype(np.float32) / 255.0
+
+                tensor5 = np.empty((5, args.out_h, args.out_w), dtype=np.float32)
+                tensor5[0] = rgb_float[:, :, 0]
+                tensor5[1] = rgb_float[:, :, 1]
+                tensor5[2] = rgb_float[:, :, 2]
+                tensor5[3] = re_float
+                tensor5[4] = nir_float
+
+                stack_ms = (time.perf_counter() - t_stack0) * 1000.0
+                stack_samples_ms.append(stack_ms)
+
+                n_triplets += 1
+
+                if args.save_debug and not saved:
+                    cv2.imwrite(str(Path(args.debug_dir) / "rgb_aligned.png"), rgb_bgr)
+                    cv2.imwrite(str(Path(args.debug_dir) / "re_aligned.png"), re_gray)
+                    cv2.imwrite(str(Path(args.debug_dir) / "nir_aligned.png"), nir_gray)
+                    np.save(str(Path(args.debug_dir) / "tensor5_sample.npy"), tensor5)
+                    saved = True
+
+                # Marca como consumidos para não recontar infinitamente o mesmo trio.
+                last_rgb = None
+                last_re = None
+                last_nir = None
+
+            if now - t_last_log >= 1.0:
+                elapsed = now - t_start
+                fps_triplets = n_triplets / max(elapsed, 1e-6)
+                fps_rgb = n_rgb / max(elapsed, 1e-6)
+                fps_re = n_re / max(elapsed, 1e-6)
+                fps_nir = n_nir / max(elapsed, 1e-6)
+
+                sync_mean = float(np.mean(lat_samples_ms[-30:])) if lat_samples_ms else -1.0
+                sync_max = float(np.max(lat_samples_ms[-30:])) if lat_samples_ms else -1.0
+                stack_mean = float(np.mean(stack_samples_ms[-30:])) if stack_samples_ms else -1.0
+                stack_max = float(np.max(stack_samples_ms[-30:])) if stack_samples_ms else -1.0
+
+                print(
+                    "[OAK_ALIGN_PERF] "
+                    f"elapsed={elapsed:.1f}s "
+                    f"triplets={n_triplets} "
+                    f"fps_triplets={fps_triplets:.2f} "
+                    f"fps_rgb={fps_rgb:.2f} "
+                    f"fps_re={fps_re:.2f} "
+                    f"fps_nir={fps_nir:.2f} "
+                    f"sync_mean_ms={sync_mean:.2f} "
+                    f"sync_max_ms={sync_max:.2f} "
+                    f"stack_mean_ms={stack_mean:.2f} "
+                    f"stack_max_ms={stack_max:.2f} "
+                    f"tensor_shape={(5, args.out_h, args.out_w)}"
+                )
+
+                t_last_log = now
+
+            time.sleep(0.001)
+
+        elapsed = time.perf_counter() - t_start
+        print("============================================")
+        print("RESULTADO FINAL")
+        print(f"elapsed       : {elapsed:.2f}s")
+        print(f"triplets      : {n_triplets}")
+        print(f"fps_triplets  : {n_triplets / max(elapsed, 1e-6):.2f}")
+        print(f"rgb frames    : {n_rgb} | fps={n_rgb / max(elapsed, 1e-6):.2f}")
+        print(f"re frames     : {n_re} | fps={n_re / max(elapsed, 1e-6):.2f}")
+        print(f"nir frames    : {n_nir} | fps={n_nir / max(elapsed, 1e-6):.2f}")
+
+        if lat_samples_ms:
+            print(f"sync mean/max : {np.mean(lat_samples_ms):.2f} / {np.max(lat_samples_ms):.2f} ms")
+
+        if stack_samples_ms:
+            print(f"stack mean/max: {np.mean(stack_samples_ms):.2f} / {np.max(stack_samples_ms):.2f} ms")
+
+        print("============================================")
+
+
+if __name__ == "__main__":
+    main()

Python/OAK/datasets/oak-fcc-3/utils/check_saved_files.py

@@ -248,8 +248,11 @@ def build_multispec_from_raw_native_multi(group: dict, meta: dict):
         if alt.exists():
             calib_path = str(alt)
         else:
-            print(f"[WARN] module_params não encontrado: {calib_path}. Tentando sem calibração.")
-            calib_path = None
+            raise RuntimeError(
+                f"module_params.json não encontrado: {calib_path}. "
+                f"Não vou gerar MULTISPEC offline sem fusion_config, "
+                f"porque isso deixaria RE/NIR desalinhados do RGB."
+            )
 
     core = RawProcessorCore(
         sensor_width=sensor_width,
@@ -271,6 +274,9 @@ def build_multispec_from_raw_native_multi(group: dict, meta: dict):
     tensor = core.build_infer_tensor_from_stream(frame, processing_meta, 5)
 
     processing_info = {
+        "fusion_result": getattr(core, "last_fusion_result", None),
+        "fusion_config_used": getattr(core, "fusion_config", None),
+        "radiometric_normalization_result": getattr(core, "last_radiometric_normalization_result", None),
         "patch_normalization_result": getattr(core, "last_patch_normalization_result", None),
         "frame_quality": getattr(core, "last_frame_quality_result", None),
     }
@@ -1135,7 +1141,7 @@ def main():
     client = OakFcc3Client(
         width=1280,
         height=800,
-        bayer="RGGB",
+        bayer="BGGR",
         frame_type="RAW_BRUTO",
         capture_mode="SINGLE",
         raw_policy="allow_single",

Python/OAK/datasets/oak-fcc-3/utils/flat_rgb_calibration_tuner.py

@@ -0,0 +1,708 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+raw_tensor_calibration_tuner.py
+
+Ferramenta visual para calibrar o tensor científico a partir do RAW_BRUTO salvo.
+
+Ela lê:
+  <grupo>/bins/<base>_CAM_A.bin
+  <grupo>/bins/<base>_CAM_B.bin
+  <grupo>/bins/<base>_CAM_C.bin
+  <grupo>/metas/<base>.json
+
+E monta o tensor final usando RawProcessorCore + module_params.json,
+passando pelo mesmo fluxo real do normalize/runtime:
+  RAW10 -> linear_demosaic/bayer_planes -> rgb_calibration
+  -> radiometric_normalization -> homography/crop/resize
+  -> flatfield final_tensor_space -> tensor [R,G,B,RE,NIR]
+
+Mostra:
+  esquerda : RGB do tensor base, com module_params original
+  centro   : RGB do tensor ajustado pelos sliders
+  direita  : diff x4 entre base e ajustado
+
+Também mostra p50/mean/p01/p99/std de cada canal do tensor ajustado.
+
+Uso:
+  python raw_tensor_calibration_tuner.py ^
+    --src_dir dataset/original/group/chao ^
+    --module_params calibration/module_params.json ^
+    --res 1024x640 ^
+    --raw_size 1280x800
+
+Controles:
+  A / seta esquerda  -> amostra anterior
+  D / seta direita   -> próxima amostra
+  S                  -> salva patch JSON
+  P                  -> imprime patch JSON
+  R                  -> reseta sliders para valores iniciais do module_params
+  Q / ESC            -> sair
+
+Notas:
+  - Sliders flat_* usam escala 1000 = strength 1.0.
+    Para flat-field, neutro é 0.
+  - Sliders rgb_gain_* usam escala 1000 = ganho 1.0.
+    Para RGB gain, neutro é 1000.
+  - auto_level/gamma afetam apenas a visualização, não o tensor.
+"""
+
+from __future__ import annotations
+
+import argparse
+import copy
+import json
+import os
+import sys
+import time
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Dict, Optional, Tuple, List
+
+import cv2
+import numpy as np
+
+
+# ============================================================
+# Import do core
+# ============================================================
+
+def add_project_paths():
+    here = Path.cwd()
+    candidates = [
+        here,
+        here / "Python" / "Scripts" / "workers" / "camera_worker" / "oak_fcc3_core",
+        here / "Scripts" / "workers" / "camera_worker" / "oak_fcc3_core",
+    ]
+
+    for c in candidates:
+        if c.exists() and str(c) not in sys.path:
+            sys.path.insert(0, str(c))
+
+add_project_paths()
+
+try:
+    from core.raw_processor_core import RawProcessorCore
+except Exception as e:
+    raise SystemExit(
+        "[ERRO] Não consegui importar core.raw_processor_core.RawProcessorCore.\n"
+        "Rode este script a partir da pasta do projeto onde existe Python/Scripts/workers/camera_worker/oak_fcc3_core,\n"
+        "ou coloque este script dentro da pasta oak_fcc3_core.\n"
+        f"Erro original: {type(e).__name__}: {e}"
+    )
+
+
+CHANNEL_NAMES = ["R", "G", "B", "RE", "NIR"]
+
+
+# ============================================================
+# IO helpers
+# ============================================================
+
+def load_json(path: Path) -> dict:
+    with path.open("r", encoding="utf-8") as f:
+        return json.load(f)
+
+
+def save_json(path: Path, data: dict):
+    path.parent.mkdir(parents=True, exist_ok=True)
+    with path.open("w", encoding="utf-8") as f:
+        json.dump(data, f, ensure_ascii=False, indent=2)
+
+
+def parse_size(txt: str) -> Tuple[int, int]:
+    w, h = [int(x) for x in str(txt).lower().split("x")]
+    return w, h
+
+
+@dataclass
+class SampleBundle:
+    base: str
+    meta_path: Path
+    bin_paths: Dict[str, Path]
+
+
+def find_sample_bins(bins_dir: Path, base: str) -> Dict[str, Path]:
+    out = {}
+    for cam_id in ("CAM_A", "CAM_B", "CAM_C"):
+        p = bins_dir / f"{base}_{cam_id}.bin"
+        if p.exists():
+            out[cam_id] = p
+    return out
+
+
+def collect_samples(src_dir: Path) -> List[SampleBundle]:
+    metas_dir = src_dir / "metas"
+    bins_dir = src_dir / "bins"
+
+    if not metas_dir.is_dir() or not bins_dir.is_dir():
+        raise RuntimeError(
+            f"src_dir precisa conter bins/ e metas/. Veio: {src_dir}"
+        )
+
+    samples = []
+    for meta_path in sorted(metas_dir.glob("*.json")):
+        base = meta_path.stem
+        bin_paths = find_sample_bins(bins_dir, base)
+
+        if not all(cam in bin_paths for cam in ("CAM_A", "CAM_B", "CAM_C")):
+            print(f"[WARN] Pulando {base}: bins incompletos {list(bin_paths.keys())}")
+            continue
+
+        samples.append(SampleBundle(base=base, meta_path=meta_path, bin_paths=bin_paths))
+
+    if not samples:
+        raise RuntimeError(f"Nenhuma amostra válida encontrada em {src_dir}")
+
+    return samples
+
+
+def load_frame_from_saved_bins(sample: SampleBundle, meta: dict) -> Dict[str, np.ndarray]:
+    saved_dtypes = meta.get("saved_payload_dtypes", {}) or {}
+    saved_shapes = meta.get("saved_payload_shapes", {}) or {}
+
+    frame = {}
+
+    for cam_id, bin_path in sample.bin_paths.items():
+        dtype = saved_dtypes.get(cam_id)
+        shape = saved_shapes.get(cam_id)
+
+        if dtype is None or shape is None:
+            raise RuntimeError(f"Faltam saved_payload_dtypes/shapes para {cam_id} em {sample.base}")
+
+        arr = np.fromfile(str(bin_path), dtype=np.dtype(dtype)).reshape(tuple(shape))
+        frame[cam_id] = arr
+
+    return frame
+
+
+def build_processing_meta(meta: dict) -> dict:
+    stream_meta = dict(meta.get("stream_meta", {}) or {})
+    stream_meta["frame_type"] = "RAW_BRUTO"
+
+    if "camera_info" not in stream_meta and isinstance(meta.get("camera_info"), dict):
+        stream_meta["camera_info"] = meta.get("camera_info")
+
+    # Mantém compatibilidade com metadados antigos e novos.
+    for key in (
+        "actual_camera_controls",
+        "startup_camera_controls",
+        "camera_controls",
+        "frame_controls",
+        "camera_frames",
+    ):
+        if meta.get(key) is not None:
+            stream_meta[key] = meta.get(key)
+
+    return stream_meta
+
+
+def tensor_stats(tensor: np.ndarray) -> Dict[str, dict]:
+    out = {}
+    for i, ch in enumerate(CHANNEL_NAMES):
+        if i >= tensor.shape[0]:
+            break
+
+        vals = tensor[i].astype(np.float32).reshape(-1)
+        vals = vals[np.isfinite(vals)]
+        if vals.size == 0:
+            out[ch] = {}
+            continue
+
+        out[ch] = {
+            "min": float(np.min(vals)),
+            "p01": float(np.percentile(vals, 1)),
+            "p50": float(np.percentile(vals, 50)),
+            "p99": float(np.percentile(vals, 99)),
+            "max": float(np.max(vals)),
+            "mean": float(np.mean(vals)),
+            "std": float(np.std(vals)),
+        }
+
+    return out
+
+
+# ============================================================
+# Config adjustment
+# ============================================================
+
+def deep_merge(dst: dict, src: dict) -> dict:
+    out = copy.deepcopy(dst)
+
+    def rec(a, b):
+        for k, v in b.items():
+            if isinstance(v, dict) and isinstance(a.get(k), dict):
+                rec(a[k], v)
+            else:
+                a[k] = copy.deepcopy(v)
+
+    rec(out, src)
+    return out
+
+
+def load_initial_values(module_params: dict):
+    ff = module_params.get("flatfield_config", {}) or {}
+    rgbc = module_params.get("rgb_calibration", {}) or {}
+
+    strengths = {ch: float(ff.get("strength", 0.0)) for ch in CHANNEL_NAMES}
+    for k, v in (ff.get("strength_by_channel", {}) or {}).items():
+        ku = str(k).upper()
+        if ku in strengths:
+            strengths[ku] = float(v)
+
+    # Contrato correto do RawProcessorCore: rgb_calibration.gains
+    gains = {"R": 1.0, "G": 1.0, "B": 1.0}
+    src = rgbc.get("gains", {}) if isinstance(rgbc, dict) else {}
+    if isinstance(src, dict):
+        for ch in ("R", "G", "B"):
+            if ch in src:
+                gains[ch] = float(src[ch])
+
+    # Compatibilidade com patch antigo do tuner.
+    src2 = rgbc.get("gain_by_channel", {}) if isinstance(rgbc, dict) else {}
+    if isinstance(src2, dict):
+        for ch in ("R", "G", "B"):
+            if ch in src2:
+                gains[ch] = float(src2[ch])
+
+    return strengths, gains
+
+
+def make_patch(strengths: Dict[str, float], rgb_gains: Dict[str, float]) -> dict:
+    return {
+        "flatfield_config": {
+            "strength_by_channel": {
+                "R": float(strengths["R"]),
+                "G": float(strengths["G"]),
+                "B": float(strengths["B"]),
+                "RE": float(strengths["RE"]),
+                "NIR": float(strengths["NIR"]),
+            }
+        },
+        "rgb_calibration": {
+            "enabled": True,
+            "gains": {
+                "R": float(rgb_gains["R"]),
+                "G": float(rgb_gains["G"]),
+                "B": float(rgb_gains["B"]),
+            }
+        }
+    }
+
+
+def apply_patch_to_core(core: RawProcessorCore, patch: dict):
+    # Flatfield.
+    ff_patch = patch.get("flatfield_config", {}) or {}
+    if isinstance(ff_patch, dict):
+        core.flatfield_config = deep_merge(core.flatfield_config, ff_patch)
+
+    # RGB calibration.
+    rgb_patch = patch.get("rgb_calibration", {}) or {}
+    if isinstance(rgb_patch, dict):
+        core.rgb_calibration = deep_merge(core.rgb_calibration, rgb_patch)
+
+    # Limpa caches cujo valor depende do flat gain/strength.
+    if hasattr(core, "_flatfield_runtime_cache"):
+        core._flatfield_runtime_cache.clear()
+
+    # Geometria/remap não depende do strength nem dos rgb gains,
+    # então não precisa limpar.
+
+
+def create_core(raw_size: Tuple[int, int], bayer_pattern: str, module_params: Path) -> RawProcessorCore:
+    raw_w, raw_h = raw_size
+    core = RawProcessorCore(
+        sensor_width=int(raw_w),
+        sensor_height=int(raw_h),
+        bayer_pattern=str(bayer_pattern or "RGGB").upper(),
+        calibration_json_path=str(module_params),
+    )
+    return core
+
+
+# ============================================================
+# Tensor generation
+# ============================================================
+
+class TensorBuilder:
+    def __init__(self, module_params: Path, raw_size: Tuple[int, int], res: Tuple[int, int]):
+        self.module_params = module_params
+        self.raw_size = raw_size
+        self.res = res
+        self.module_params_data = load_json(module_params)
+
+        self.base_core = None
+        self.tune_core = None
+        self.current_bayer = None
+
+        self._sample_cache = {}
+        self._base_tensor_cache = {}
+
+    def ensure_cores(self, bayer_pattern: str):
+        bayer = str(bayer_pattern or "RGGB").upper()
+
+        if self.base_core is not None and self.current_bayer == bayer:
+            return
+
+        print(f"[CORE] criando cores bayer={bayer} raw={self.raw_size} res={self.res}")
+        self.base_core = create_core(self.raw_size, bayer, self.module_params)
+        self.tune_core = create_core(self.raw_size, bayer, self.module_params)
+        self.current_bayer = bayer
+
+    def get_frame_meta(self, sample: SampleBundle):
+        key = sample.base
+        cached = self._sample_cache.get(key)
+        if cached is not None:
+            return cached
+
+        meta = load_json(sample.meta_path)
+        frame = load_frame_from_saved_bins(sample, meta)
+        processing_meta = build_processing_meta(meta)
+        bayer = str(meta.get("bayer_pattern") or meta.get("bayer") or "RGGB").upper()
+
+        cached = (frame, processing_meta, bayer, meta)
+        self._sample_cache[key] = cached
+        return cached
+
+    def build_base_tensor(self, sample: SampleBundle):
+        if sample.base in self._base_tensor_cache:
+            return self._base_tensor_cache[sample.base]
+
+        frame, processing_meta, bayer, _meta = self.get_frame_meta(sample)
+        self.ensure_cores(bayer)
+
+        tensor = self.base_core.build_infer_tensor_from_stream(
+            frame=frame,
+            meta=processing_meta,
+            channels_expected=5,
+            target_size=self.res,
+        )
+
+        tensor = np.ascontiguousarray(tensor.astype(np.float32, copy=False))
+        self._base_tensor_cache[sample.base] = tensor
+        return tensor
+
+    def build_tuned_tensor(self, sample: SampleBundle, strengths: Dict[str, float], rgb_gains: Dict[str, float]):
+        frame, processing_meta, bayer, _meta = self.get_frame_meta(sample)
+        self.ensure_cores(bayer)
+
+        patch = make_patch(strengths, rgb_gains)
+        apply_patch_to_core(self.tune_core, patch)
+
+        tensor = self.tune_core.build_infer_tensor_from_stream(
+            frame=frame,
+            meta=processing_meta,
+            channels_expected=5,
+            target_size=self.res,
+        )
+
+        return np.ascontiguousarray(tensor.astype(np.float32, copy=False))
+
+
+# ============================================================
+# Display
+# ============================================================
+
+def val_to_gain(v: int) -> float:
+    return float(v) / 1000.0
+
+
+def gain_to_val(g: float) -> int:
+    return int(round(max(0.0, min(3.0, float(g))) * 1000.0))
+
+
+def val_to_strength(v: int) -> float:
+    return float(v) / 1000.0
+
+
+def strength_to_val(s: float) -> int:
+    return int(round(max(0.0, min(2.0, float(s))) * 1000.0))
+
+
+def read_ui(win: str):
+    strengths = {
+        "R": val_to_strength(cv2.getTrackbarPos("flat_R", win)),
+        "G": val_to_strength(cv2.getTrackbarPos("flat_G", win)),
+        "B": val_to_strength(cv2.getTrackbarPos("flat_B", win)),
+        "RE": val_to_strength(cv2.getTrackbarPos("flat_RE", win)),
+        "NIR": val_to_strength(cv2.getTrackbarPos("flat_NIR", win)),
+    }
+
+    gains = {
+        "R": val_to_gain(cv2.getTrackbarPos("rgb_gain_R", win)),
+        "G": val_to_gain(cv2.getTrackbarPos("rgb_gain_G", win)),
+        "B": val_to_gain(cv2.getTrackbarPos("rgb_gain_B", win)),
+    }
+
+    auto_level = bool(cv2.getTrackbarPos("auto_level", win))
+    gamma = max(0.05, cv2.getTrackbarPos("gamma_x100", win) / 100.0)
+
+    return strengths, gains, auto_level, gamma
+
+
+def set_ui(win: str, strengths: Dict[str, float], gains: Dict[str, float], gamma: float):
+    for ch in CHANNEL_NAMES:
+        cv2.setTrackbarPos(f"flat_{ch}", win, strength_to_val(strengths.get(ch, 0.0)))
+    for ch in ("R", "G", "B"):
+        cv2.setTrackbarPos(f"rgb_gain_{ch}", win, gain_to_val(gains.get(ch, 1.0)))
+    cv2.setTrackbarPos("gamma_x100", win, int(round(gamma * 100)))
+
+
+def preview_rgb_from_tensor(tensor: np.ndarray, auto_level: bool, gamma: float) -> np.ndarray:
+    rgb = np.transpose(tensor[:3].astype(np.float32), (1, 2, 0))
+    rgb = np.nan_to_num(rgb, nan=0.0, posinf=1.0, neginf=0.0)
+
+    if auto_level:
+        # Auto-level visual separado por canal para facilitar inspeção.
+        out = np.empty_like(rgb)
+        for c in range(3):
+            ch = rgb[..., c]
+            lo = float(np.percentile(ch, 1.0))
+            hi = float(np.percentile(ch, 99.0))
+            if hi > lo:
+                out[..., c] = (ch - lo) / (hi - lo)
+            else:
+                out[..., c] = ch
+        rgb = out
+
+    rgb = np.clip(rgb, 0.0, 1.0)
+
+    if gamma > 0 and abs(gamma - 1.0) > 1e-6:
+        rgb = np.power(rgb, gamma)
+
+    return (rgb * 255.0).clip(0, 255).astype(np.uint8)
+
+
+def put_text(img: np.ndarray, lines, x=12, y=25, scale=0.55):
+    out = img.copy()
+    yy = y
+    for line in lines:
+        cv2.putText(out, str(line), (x, yy), cv2.FONT_HERSHEY_SIMPLEX, scale, (0, 0, 0), 4, cv2.LINE_AA)
+        cv2.putText(out, str(line), (x, yy), cv2.FONT_HERSHEY_SIMPLEX, scale, (255, 255, 255), 1, cv2.LINE_AA)
+        yy += int(22 * scale / 0.55)
+    return out
+
+
+def resize_keep_width(img: np.ndarray, target_w: int) -> np.ndarray:
+    h, w = img.shape[:2]
+    if w <= target_w:
+        return img
+    scale = target_w / float(w)
+    return cv2.resize(img, (target_w, int(h * scale)), interpolation=cv2.INTER_AREA)
+
+
+def stats_lines(stats: Dict[str, dict], prefix=""):
+    lines = []
+    for ch in CHANNEL_NAMES:
+        s = stats.get(ch, {}) or {}
+        if not s:
+            continue
+        lines.append(
+            f"{prefix}{ch}: p50={s['p50']:.4f} mean={s['mean']:.4f} "
+            f"p01={s['p01']:.4f} p99={s['p99']:.4f} std={s['std']:.4f}"
+        )
+    return lines
+
+
+def build_canvas(
+    sample: SampleBundle,
+    idx: int,
+    total: int,
+    base_tensor: np.ndarray,
+    tuned_tensor: np.ndarray,
+    strengths: Dict[str, float],
+    gains: Dict[str, float],
+    auto_level: bool,
+    gamma: float,
+    max_width: int,
+):
+    base_rgb = preview_rgb_from_tensor(base_tensor, auto_level=auto_level, gamma=gamma)
+    tuned_rgb = preview_rgb_from_tensor(tuned_tensor, auto_level=auto_level, gamma=gamma)
+
+    diff = np.abs(tuned_rgb.astype(np.int16) - base_rgb.astype(np.int16)).clip(0, 255).astype(np.uint8)
+    diff = np.clip(diff.astype(np.uint16) * 4, 0, 255).astype(np.uint8)
+
+    h = base_rgb.shape[0]
+    sep = np.zeros((h, 4, 3), dtype=np.uint8)
+    canvas = np.hstack([base_rgb, sep, tuned_rgb, sep, diff])
+
+    st_base = tensor_stats(base_tensor)
+    st_tuned = tensor_stats(tuned_tensor)
+
+    top_lines = [
+        f"{idx + 1}/{total} | {sample.base}",
+        "LEFT=base tensor | MID=tuned tensor | RIGHT=diff x4",
+        f"flat R={strengths['R']:.3f} G={strengths['G']:.3f} B={strengths['B']:.3f} RE={strengths['RE']:.3f} NIR={strengths['NIR']:.3f}",
+        f"rgb_gain R={gains['R']:.3f} G={gains['G']:.3f} B={gains['B']:.3f} | auto={int(auto_level)} gamma={gamma:.2f}",
+        "A/D navega | S salva patch | P print | R reset | Q sai",
+    ]
+
+    canvas = put_text(canvas, top_lines, x=12, y=24, scale=0.56)
+
+    # Painel de stats na parte inferior, com fundo escuro.
+    panel_h = 190
+    panel = np.zeros((panel_h, canvas.shape[1], 3), dtype=np.uint8)
+
+    stat_lines = ["TUNED tensor stats:"] + stats_lines(st_tuned, "")
+    # Também mostra relação G/R e G/B, útil para diagnosticar esverdeado.
+    try:
+        gr = st_tuned["G"]["mean"] / max(st_tuned["R"]["mean"], 1e-6)
+        gb = st_tuned["G"]["mean"] / max(st_tuned["B"]["mean"], 1e-6)
+        stat_lines.append(f"RGB balance mean: G/R={gr:.3f} | G/B={gb:.3f}")
+    except Exception:
+        pass
+
+    panel = put_text(panel, stat_lines[:8], x=12, y=24, scale=0.58)
+
+    full = np.vstack([canvas, panel])
+    return resize_keep_width(full, max_width)
+
+
+# ============================================================
+# Main
+# ============================================================
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--src_dir", required=True, help="Grupo original contendo bins/ e metas/. Ex: dataset/original/group/chao")
+    ap.add_argument("--module_params", required=True, help="calibration/module_params.json")
+    ap.add_argument("--res", default="1024x640", help="Resolução final WxH do tensor")
+    ap.add_argument("--raw_size", default="1280x800", help="RAW size WxH")
+    ap.add_argument("--out_patch", default="raw_tensor_tuned_patch.json")
+    ap.add_argument("--max_width", type=int, default=1900)
+    ap.add_argument("--gamma", type=float, default=1.0)
+    args = ap.parse_args()
+
+    src_dir = Path(args.src_dir)
+    module_params = Path(args.module_params)
+
+    if not src_dir.is_dir():
+        raise SystemExit(f"[ERRO] src_dir não existe: {src_dir}")
+
+    if not module_params.is_file():
+        raise SystemExit(f"[ERRO] module_params não existe: {module_params}")
+
+    res = parse_size(args.res)
+    raw_size = parse_size(args.raw_size)
+
+    module_data = load_json(module_params)
+    initial_strengths, initial_gains = load_initial_values(module_data)
+
+    samples = collect_samples(src_dir)
+
+    builder = TensorBuilder(
+        module_params=module_params,
+        raw_size=raw_size,
+        res=res,
+    )
+
+    win = "RAW Tensor Calibration Tuner"
+    cv2.namedWindow(win, cv2.WINDOW_NORMAL)
+
+    for ch in CHANNEL_NAMES:
+        cv2.createTrackbar(f"flat_{ch}", win, 0, 2000, lambda v: None)
+
+    for ch in ("R", "G", "B"):
+        cv2.createTrackbar(f"rgb_gain_{ch}", win, 1000, 3000, lambda v: None)
+
+    cv2.createTrackbar("auto_level", win, 0, 1, lambda v: None)
+    cv2.createTrackbar("gamma_x100", win, 100, 300, lambda v: None)
+
+    set_ui(win, initial_strengths, initial_gains, args.gamma)
+
+    idx = 0
+    last_key = None
+    last_tuned_key = None
+    base_tensor = None
+    tuned_tensor = None
+
+    print("[INFO] RAW Tensor Calibration Tuner iniciado.")
+    print("[INFO] O tensor base usa module_params original. O tensor tuned usa sliders.")
+    print("[INFO] S salva patch JSON. P imprime valores. R reseta.")
+    print("[INFO] Se ficar lento ao arrastar sliders, solte o mouse: ele recalcula o tensor real a cada mudança.")
+
+    while True:
+        sample = samples[idx]
+
+        strengths, gains, auto_level, gamma = read_ui(win)
+
+        tuned_key = (
+            sample.base,
+            tuple(round(strengths[ch], 4) for ch in CHANNEL_NAMES),
+            tuple(round(gains[ch], 4) for ch in ("R", "G", "B")),
+        )
+
+        try:
+            if last_key != sample.base:
+                t0 = time.perf_counter()
+                base_tensor = builder.build_base_tensor(sample)
+                tuned_tensor = None
+                last_tuned_key = None
+                last_key = sample.base
+                print(f"[LOAD] {sample.base} base_tensor={base_tensor.shape} {((time.perf_counter()-t0)*1000):.1f}ms")
+
+            if last_tuned_key != tuned_key or tuned_tensor is None:
+                t0 = time.perf_counter()
+                tuned_tensor = builder.build_tuned_tensor(sample, strengths, gains)
+                last_tuned_key = tuned_key
+                print(f"[TUNE] {sample.base} tuned_tensor={tuned_tensor.shape} {((time.perf_counter()-t0)*1000):.1f}ms")
+
+            canvas = build_canvas(
+                sample=sample,
+                idx=idx,
+                total=len(samples),
+                base_tensor=base_tensor,
+                tuned_tensor=tuned_tensor,
+                strengths=strengths,
+                gains=gains,
+                auto_level=auto_level,
+                gamma=gamma,
+                max_width=args.max_width,
+            )
+
+            cv2.imshow(win, cv2.cvtColor(canvas, cv2.COLOR_RGB2BGR))
+
+        except Exception as e:
+            err = np.zeros((600, 1200, 3), dtype=np.uint8)
+            err = put_text(err, [
+                f"ERRO em {sample.base}",
+                f"{type(e).__name__}: {e}",
+                "A/D navega | Q sai",
+            ], x=20, y=40, scale=0.7)
+            cv2.imshow(win, cv2.cvtColor(err, cv2.COLOR_RGB2BGR))
+
+        k = cv2.waitKey(30) & 0xFF
+
+        if k in (ord("q"), ord("Q"), 27):
+            break
+
+        elif k in (ord("d"), ord("D"), 83):
+            idx = (idx + 1) % len(samples)
+            last_key = None
+            last_tuned_key = None
+
+        elif k in (ord("a"), ord("A"), 81):
+            idx = (idx - 1 + len(samples)) % len(samples)
+            last_key = None
+            last_tuned_key = None
+
+        elif k in (ord("r"), ord("R")):
+            set_ui(win, initial_strengths, initial_gains, args.gamma)
+            last_tuned_key = None
+
+        elif k in (ord("p"), ord("P")):
+            patch = make_patch(strengths, gains)
+            print(json.dumps(patch, ensure_ascii=False, indent=2))
+
+        elif k in (ord("s"), ord("S")):
+            patch = make_patch(strengths, gains)
+            out = Path(args.out_patch)
+            save_json(out, patch)
+            print(f"[SAVE] {out.resolve()}")
+            print(json.dumps(patch, ensure_ascii=False, indent=2))
+
+    cv2.destroyAllWindows()
+
+
+if __name__ == "__main__":
+    main()

Python/OAK/datasets/oak-fcc-3/utils/manual_fusion_calibrator.py

@@ -635,6 +635,7 @@ def main():
                     offsets_data.setdefault("homographies", {})
                     offsets_data["schema"] = "manual_multispec_offsets_v2"
                     offsets_data["reference_camera"] = "rgb"
+                    offsets_data["homography_calibration_size"] = [int(base_w), int(base_h)]
 
                     save_offsets_json(args.out_json, offsets_data)
                     last_msg = f"Offsets salvos em: {args.out_json}"

Python/raspi/Instalar Sistema.txt

@@ -0,0 +1,92 @@
+1. Conectar cartão SD no adaptador e espetar na porta USB
+
+2. Baixar ferramenta de instalação do site
+https://www.raspberrypi.com/software/
+
+3. Instalar ferramenta
+
+4. Abrir
+
+5. Configurar Sistema Headless
+Choose Device → Raspberry Pi 5
+Choose OS → Raspberry Pi OS Other → Raspberry Pi OS Lite 64-bit
+Choose Storage → selecione o cartão SD
+Clique em Next ou configurações de personalização
+Hostname: pi-cam
+Usuário: diego
+Senha: 1234
+Enable SSH → Use password authentication
+Country: Brasília - Brasil
+Timezone: America/Sao_Paulo
+Keyboard: br
+Write
+
+6. Remover SD do computador
+
+7. Colocar o SD no Pi
+
+8. Conectar cabo de rede entre Pi e PC
+
+9. Conectar a fonte
+
+10. Espere uns 2 a 5 minutos no primeiro boot
+
+11. Tentar conectar via SSH
+ssh diego@pi-cam.local
+
+Se encontrar, digitar yes, e em seguida a senha do usuário para acessar: 1234
+
+12. Atualizar Sistema
+sudo apt update
+sudo apt full-upgrade -y
+sudo reboot
+
+11. Validar configuração do Pi
+sudo raspi-config
+
+Interface Options → SSH → Enable
+Localisation Options → Timezone → America/Sao_Paulo
+System Options → Hostname → pi-cam
+
+12. Conferencias finais
+uname -a
+cat /etc/os-release
+vcgencmd measure_temp
+
+13. Definir IP Fixo (opcional)
+sudo nmcli connection modify netplan-eth0 ipv4.addresses 192.168.105.6/24 ipv4.method manual
+sudo nmcli connection up netplan-eth0
+sudo nmcli device connect eth0
+
+13. Instalar dependencias dev
+sudo apt install -y python3 python3-pip python3-venv git
+
+mkdir ~/multispec_module
+cd ~/multispec_module
+python3 -m venv venv
+source venv/bin/activate
+
+pip install numpy opencv-python
+
+sudo apt install -y python3-picamera2 libcamera-apps v4l-utils
+
+14. Conectar no VSCode através do SSH
+No VSCode, instalar a extensão Remote - SSH
+Apertar Ctrl + Shift + P
+Remote-SSH: Connect to Host
++ Add New SSH Host
+ssh diego@pi-cam.local
+
+Salvar no config padrão
+
+Ctrl + Shift + P
+→ Remote-SSH: Connect to Host
+→ pi5-local
+
+Digitar a senha
+
+File → Open Folder
+/home/diego/multispec_module
+
+Digitar a senha
+

Python/raspi/cam_3/multispectral_client.py

@@ -216,7 +216,23 @@ class MultiSpectralClient:
 
     def get_next_decoded(self, timeout=2.0, update_radiometry=True):
         frame, meta = self.get_next_frame(timeout=timeout)
-        decoded = self.core.decode_stream_cameras(frame, meta)
+
+        frame_type = meta.get("frame_type")
+
+        if frame_type == "RAW_BRUTO":
+            decoded = self.core.decode_stream_cameras(frame, meta)
+
+        elif frame_type in ("RGB", "MULTISPEC"):
+            decoded = {
+                "cam2": {
+                    "name": "RGB",
+                    "image": frame,
+                    "meta": meta
+                }
+            }
+
+        else:
+            raise RuntimeError(f"frame_type não suportado: {frame_type}")
 
         if update_radiometry:
             self.update_radiometry(decoded, meta)