Add ROI-based shoe training workflow

02_train.bat

@@ -2,8 +2,34 @@
 chcp 65001 >nul
 cls
 
-:: 设置 Python 3.11 路径
+set "REPO_DIR=%~dp0"
-set "PATH=C:\Users\Tellme\AppData\Local\Programs\Python\Python311\Scripts;C:\Users\Tellme\AppData\Local\Programs\Python\Python311;%PATH%"
+pushd "%REPO_DIR%"
+
+set "YOLO_LAUNCHER="
+where yolo >nul 2>nul
+if %ERRORLEVEL% equ 0 set "YOLO_LAUNCHER=yolo"
+
+if not defined YOLO_LAUNCHER (
+    py -3.11 -c "import ultralytics" >nul 2>nul
+    if %ERRORLEVEL% equ 0 set "YOLO_LAUNCHER=py -3.11 -m ultralytics"
+)
+
+if not defined YOLO_LAUNCHER (
+    python -c "import ultralytics" >nul 2>nul
+    if %ERRORLEVEL% equ 0 set "YOLO_LAUNCHER=python -m ultralytics"
+)
+
+if not defined YOLO_LAUNCHER (
+    echo [错误] 未找到可用的 Ultralytics 启动方式
+    echo.
+    echo 请先确保满足以下任一条件:
+    echo   1. yolo 命令已加入 PATH
+    echo   2. py -3.11 可运行并已安装 ultralytics
+    echo   3. python 可运行并已安装 ultralytics
+    popd
+    pause
+    exit /b 1
+)
 
 echo ============================================================
 echo           训练鞋子检测模型 (YOLOv8 + 640x640)
@@ -11,14 +37,15 @@ echo ============================================================
 echo.
 
 :: 设置数据集路径
-set DATASET=datasets/openimages-shoes-yolo/data.yaml
+set "DATASET=%REPO_DIR%datasets\openimages-shoes-yolo\data.yaml"
 
 :: 检查数据集是否存在
-if not exist %DATASET% (
+if not exist "%DATASET%" (
     echo [错误] 找不到数据集配置文件: %DATASET%
     echo.
     echo 请先下载数据集:
-    echo   python 01_download_dataset.py --source openimages --max-samples 5000
+    echo   py -3.11 "%REPO_DIR%01_download_dataset.py" --source openimages --max-samples 5000
+    popd
     pause
     exit /b 1
 )
@@ -69,11 +96,12 @@ echo                     开始训练
 echo ============================================================
 echo.
 
-yolo detect train data=%DATASET% model=%MODEL% epochs=%EPOCHS% imgsz=%IMGSZ% batch=%BATCH% device=0
+call %YOLO_LAUNCHER% detect train data="%DATASET%" model="%MODEL%" epochs=%EPOCHS% imgsz=%IMGSZ% batch=%BATCH% device=0
 
 if %ERRORLEVEL% neq 0 (
     echo.
     echo [错误] 训练失败！
+    popd
     pause
     exit /b 1
 )
@@ -87,4 +115,5 @@ echo 模型保存在: runs/detect/train/weights/best.pt
 echo.
 echo 下一步: 运行 03_export_onnx.bat 导出 ONNX
 echo.
+popd
 pause

03_export_onnx.bat

@@ -2,19 +2,46 @@
 chcp 65001 >nul
 cls
 
-:: 设置 Python 3.11 路径
+set "REPO_DIR=%~dp0"
-set "PATH=C:\Users\Tellme\AppData\Local\Programs\Python\Python311\Scripts;C:\Users\Tellme\AppData\Local\Programs\Python\Python311;%PATH%"
+pushd "%REPO_DIR%"
+
+set "YOLO_LAUNCHER="
+where yolo >nul 2>nul
+if %ERRORLEVEL% equ 0 set "YOLO_LAUNCHER=yolo"
+
+if not defined YOLO_LAUNCHER (
+    py -3.11 -c "import ultralytics" >nul 2>nul
+    if %ERRORLEVEL% equ 0 set "YOLO_LAUNCHER=py -3.11 -m ultralytics"
+)
+
+if not defined YOLO_LAUNCHER (
+    python -c "import ultralytics" >nul 2>nul
+    if %ERRORLEVEL% equ 0 set "YOLO_LAUNCHER=python -m ultralytics"
+)
+
+if not defined YOLO_LAUNCHER (
+    echo [错误] 未找到可用的 Ultralytics 启动方式
+    echo.
+    echo 请先确保满足以下任一条件:
+    echo   1. yolo 命令已加入 PATH
+    echo   2. py -3.11 可运行并已安装 ultralytics
+    echo   3. python 可运行并已安装 ultralytics
+    popd
+    pause
+    exit /b 1
+)
 
 echo ============================================================
 echo         导出 ONNX 模型 (640x640)
 echo ============================================================
 echo.
 
-set MODEL_PATH=%USERPROFILE%\apps\ultralytics\runs\detect\train\weights\best.pt
+set "MODEL_PATH=%REPO_DIR%runs\detect\train\weights\best.pt"
 
-if not exist %MODEL_PATH% (
+if not exist "%MODEL_PATH%" (
     echo [错误] 找不到模型: %MODEL_PATH%
     echo 请先运行 02_train.bat 训练
+    popd
     pause
     exit /b 1
 )
@@ -22,17 +49,19 @@ if not exist %MODEL_PATH% (
 echo [信息] 输入模型: %MODEL_PATH%
 echo.
 
-yolo export model=%MODEL_PATH% format=onnx imgsz=640 opset=12 simplify
+call %YOLO_LAUNCHER% export model="%MODEL_PATH%" format=onnx imgsz=640 opset=12 simplify
 
 if %ERRORLEVEL% neq 0 (
     echo [错误] 导出失败！
+    popd
     pause
     exit /b 1
 )
 
 echo.
-echo [成功] ONNX 模型: %USERPROFILE%\apps\ultralytics\runs\detect\train\weights\best.onnx
+echo [成功] ONNX 模型: %REPO_DIR%runs\detect\train\weights\best.onnx
 echo.
 echo 下一步: 在 Ubuntu 上运行 04_convert_rknn.py 转换
 echo.
+popd
 pause

06_finetune_ppe.bat

@@ -1,26 +1,56 @@
 @echo off
 setlocal
 
-set "PATH=C:\Users\Tellme\AppData\Local\Programs\Python\Python311\Scripts;C:\Users\Tellme\AppData\Local\Programs\Python\Python311;%PATH%"
+set "REPO_DIR=%~dp0"
+pushd "%REPO_DIR%"
+
+set "YOLO_LAUNCHER="
+where yolo >nul 2>nul
+if %ERRORLEVEL% equ 0 set "YOLO_LAUNCHER=yolo"
+
+if not defined YOLO_LAUNCHER (
+    py -3.11 -c "import ultralytics" >nul 2>nul
+    if %ERRORLEVEL% equ 0 set "YOLO_LAUNCHER=py -3.11 -m ultralytics"
+)
+
+if not defined YOLO_LAUNCHER (
+    python -c "import ultralytics" >nul 2>nul
+    if %ERRORLEVEL% equ 0 set "YOLO_LAUNCHER=python -m ultralytics"
+)
+
+if not defined YOLO_LAUNCHER (
+    echo [ERROR] No usable Ultralytics launcher was found
+    echo.
+    echo Please make sure one of the following works:
+    echo   1. yolo in PATH
+    echo   2. py -3.11 with ultralytics installed
+    echo   3. python with ultralytics installed
+    popd
+    pause
+    exit /b 1
+)
 
 echo ============================================================
 echo        Stage 2 Fine-tuning on PPE shoe subset
 echo ============================================================
 echo.
 
-set "DATASET=C:\Users\Tellme\apps\ppe-model-training\datasets\ppe-shoes\data.yaml"
+set "DATASET=%REPO_DIR%datasets\ppe-shoes\data.yaml"
-set "BASE_MODEL=C:\Users\Tellme\apps\ultralytics\runs\detect\train3\weights\best.pt"
+set "BASE_MODEL=%REPO_DIR%runs\detect\train\weights\best.pt"
 
 if not exist "%DATASET%" (
     echo [ERROR] PPE shoe subset not found: %DATASET%
     echo Run:
-    echo   C:\Users\Tellme\AppData\Local\Programs\Python\Python311\python.exe C:\Users\Tellme\apps\ppe-model-training\05_prepare_ppe_shoe_subset.py
+    echo   python "%REPO_DIR%05_prepare_ppe_shoe_subset.py"
+    popd
     pause
     exit /b 1
 )
 
 if not exist "%BASE_MODEL%" (
     echo [ERROR] Base model not found: %BASE_MODEL%
+    echo Run 02_train.bat first to create the base checkpoint.
+    popd
     pause
     exit /b 1
 )
@@ -45,11 +75,12 @@ echo                    Start fine-tuning
 echo ============================================================
 echo.
 
-yolo detect train data="%DATASET%" model="%BASE_MODEL%" epochs=%EPOCHS% imgsz=%IMGSZ% batch=%BATCH% device=0
+call %YOLO_LAUNCHER% detect train data="%DATASET%" model="%BASE_MODEL%" epochs=%EPOCHS% imgsz=%IMGSZ% batch=%BATCH% device=0
 
 if %ERRORLEVEL% neq 0 (
     echo.
     echo [ERROR] Fine-tuning failed.
+    popd
     pause
     exit /b 1
 )
@@ -59,4 +90,5 @@ echo ============================================================
 echo                    Fine-tuning complete
 echo ============================================================
 echo.
+popd
 pause

09_build_roi_shoe_dataset.py

@@ -0,0 +1,456 @@
+#!/usr/bin/env python3
+"""Build a foot-ROI shoe dataset from existing YOLO shoe datasets."""
+
+from __future__ import annotations
+
+import argparse
+import math
+import shutil
+from collections import defaultdict
+from dataclasses import dataclass
+from pathlib import Path
+
+from PIL import Image
+
+
+DEFAULT_SOURCES = [
+    "datasets/openimages-shoes-yolo",
+    "datasets/ppe-shoes",
+]
+
+IMAGE_EXTS = (".jpg", ".jpeg", ".png", ".bmp", ".webp")
+PAIR_MAX_X_GAP_FACTOR = 3.2
+PAIR_MAX_Y_GAP_FACTOR = 1.2
+PAIR_MIN_AREA_RATIO = 0.4
+PAIR_MAX_AREA_RATIO = 2.5
+
+SINGLE_AREA_RANGE = (0.15, 0.35)
+PAIR_AREA_RANGE = (0.25, 0.50)
+
+
+@dataclass(frozen=True)
+class Box:
+    x1: float
+    y1: float
+    x2: float
+    y2: float
+
+    @property
+    def w(self) -> float:
+        return max(0.0, self.x2 - self.x1)
+
+    @property
+    def h(self) -> float:
+        return max(0.0, self.y2 - self.y1)
+
+    @property
+    def area(self) -> float:
+        return self.w * self.h
+
+    @property
+    def cx(self) -> float:
+        return (self.x1 + self.x2) / 2.0
+
+    @property
+    def cy(self) -> float:
+        return (self.y1 + self.y2) / 2.0
+
+    def clip(self, width: float, height: float) -> "Box | None":
+        x1 = min(max(self.x1, 0.0), width)
+        y1 = min(max(self.y1, 0.0), height)
+        x2 = min(max(self.x2, 0.0), width)
+        y2 = min(max(self.y2, 0.0), height)
+        if x2 <= x1 or y2 <= y1:
+            return None
+        return Box(x1, y1, x2, y2)
+
+
+@dataclass(frozen=True)
+class RoiSample:
+    roi: Box
+    members: tuple[int, ...]
+    mode: str
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Build a foot-context ROI shoe dataset")
+    parser.add_argument(
+        "--sources",
+        nargs="+",
+        default=DEFAULT_SOURCES,
+        help="Source YOLO datasets containing images/<split> and labels/<split>",
+    )
+    parser.add_argument(
+        "--output",
+        default="datasets/shoe-roi-mix",
+        help="Output ROI dataset directory",
+    )
+    parser.add_argument(
+        "--clean",
+        action="store_true",
+        help="Delete the output directory before rebuilding",
+    )
+    return parser.parse_args()
+
+
+def ensure_output_layout(output_dir: Path) -> None:
+    for split in ("train", "val", "test"):
+        (output_dir / "images" / split).mkdir(parents=True, exist_ok=True)
+        (output_dir / "labels" / split).mkdir(parents=True, exist_ok=True)
+
+
+def find_image(image_dir: Path, stem: str) -> Path | None:
+    for ext in IMAGE_EXTS:
+        candidate = image_dir / f"{stem}{ext}"
+        if candidate.exists():
+            return candidate
+    return None
+
+
+def load_boxes(label_path: Path, image_width: int, image_height: int) -> list[Box]:
+    boxes: list[Box] = []
+    for raw_line in label_path.read_text(encoding="utf-8").splitlines():
+        line = raw_line.strip()
+        if not line:
+            continue
+        parts = line.split()
+        if len(parts) < 5:
+            continue
+        _, xc, yc, w, h = parts[:5]
+        box_w = float(w) * image_width
+        box_h = float(h) * image_height
+        center_x = float(xc) * image_width
+        center_y = float(yc) * image_height
+        box = Box(
+            center_x - box_w / 2.0,
+            center_y - box_h / 2.0,
+            center_x + box_w / 2.0,
+            center_y + box_h / 2.0,
+        ).clip(image_width, image_height)
+        if box is not None and box.area > 1.0:
+            boxes.append(box)
+    return dedupe_boxes(boxes)
+
+
+def dedupe_boxes(boxes: list[Box], iou_threshold: float = 0.9) -> list[Box]:
+    kept: list[Box] = []
+    for box in sorted(boxes, key=lambda item: item.area, reverse=True):
+        if any(iou(box, existing) >= iou_threshold for existing in kept):
+            continue
+        kept.append(box)
+    return sorted(kept, key=lambda item: (item.cx, item.cy))
+
+
+def iou(a: Box, b: Box) -> float:
+    inter_x1 = max(a.x1, b.x1)
+    inter_y1 = max(a.y1, b.y1)
+    inter_x2 = min(a.x2, b.x2)
+    inter_y2 = min(a.y2, b.y2)
+    inter_w = max(0.0, inter_x2 - inter_x1)
+    inter_h = max(0.0, inter_y2 - inter_y1)
+    inter_area = inter_w * inter_h
+    if inter_area <= 0:
+        return 0.0
+    union = a.area + b.area - inter_area
+    return inter_area / union if union > 0 else 0.0
+
+
+def should_pair(left: Box, right: Box) -> bool:
+    width_ref = max(left.w, right.w)
+    height_ref = max(left.h, right.h)
+    if width_ref <= 0 or height_ref <= 0:
+        return False
+
+    dx = abs(left.cx - right.cx)
+    dy = abs(left.cy - right.cy)
+    area_ratio = left.area / right.area if right.area > 0 else math.inf
+
+    return (
+        dx <= width_ref * PAIR_MAX_X_GAP_FACTOR
+        and dy <= height_ref * PAIR_MAX_Y_GAP_FACTOR
+        and PAIR_MIN_AREA_RATIO <= area_ratio <= PAIR_MAX_AREA_RATIO
+    )
+
+
+def greedy_group_boxes(boxes: list[Box]) -> list[tuple[int, ...]]:
+    if len(boxes) < 2:
+        return [(idx,) for idx in range(len(boxes))]
+
+    candidates: list[tuple[float, int, int]] = []
+    for i in range(len(boxes)):
+        for j in range(i + 1, len(boxes)):
+            if not should_pair(boxes[i], boxes[j]):
+                continue
+            dx = abs(boxes[i].cx - boxes[j].cx)
+            dy = abs(boxes[i].cy - boxes[j].cy)
+            score = dx + (0.5 * dy)
+            candidates.append((score, i, j))
+
+    used: set[int] = set()
+    groups: list[tuple[int, ...]] = []
+    for _, i, j in sorted(candidates, key=lambda item: item[0]):
+        if i in used or j in used:
+            continue
+        used.add(i)
+        used.add(j)
+        groups.append((i, j))
+
+    for idx in range(len(boxes)):
+        if idx not in used:
+            groups.append((idx,))
+
+    return groups
+
+
+def expand_single(box: Box) -> Box:
+    return Box(
+        box.x1 - (0.6 * box.w),
+        box.y1 - (0.5 * box.h),
+        box.x1 - (0.6 * box.w) + (2.2 * box.w),
+        box.y1 - (0.5 * box.h) + (2.4 * box.h),
+    )
+
+
+def expand_pair(boxes: list[Box], group: tuple[int, int]) -> Box:
+    first = boxes[group[0]]
+    second = boxes[group[1]]
+    union_x1 = min(first.x1, second.x1)
+    union_y1 = min(first.y1, second.y1)
+    union_x2 = max(first.x2, second.x2)
+    union_y2 = max(first.y2, second.y2)
+    union_w = union_x2 - union_x1
+    union_h = union_y2 - union_y1
+    roi_x = union_x1 - (0.35 * union_w)
+    roi_y = union_y1 - (0.45 * union_h)
+    return Box(
+        roi_x,
+        roi_y,
+        roi_x + (1.7 * union_w),
+        roi_y + (2.0 * union_h),
+    )
+
+
+def clamp_roi(roi: Box, image_width: int, image_height: int) -> Box | None:
+    clipped = roi.clip(float(image_width), float(image_height))
+    if clipped is None:
+        return None
+
+    x1 = int(math.floor(clipped.x1))
+    y1 = int(math.floor(clipped.y1))
+    x2 = int(math.ceil(clipped.x2))
+    y2 = int(math.ceil(clipped.y2))
+
+    x1 = max(0, min(x1, image_width - 1))
+    y1 = max(0, min(y1, image_height - 1))
+    x2 = max(x1 + 1, min(x2, image_width))
+    y2 = max(y1 + 1, min(y2, image_height))
+    return Box(float(x1), float(y1), float(x2), float(y2))
+
+
+def resize_roi_to_ratio(
+    roi: Box,
+    image_width: int,
+    image_height: int,
+    object_area: float,
+    min_ratio: float,
+    max_ratio: float,
+) -> Box | None:
+    if object_area <= 0:
+        return None
+
+    adjusted = roi
+    target_ratio = (min_ratio + max_ratio) / 2.0
+    for _ in range(3):
+        roi_area = adjusted.area
+        if roi_area <= 0:
+            return None
+        ratio = object_area / roi_area
+        if min_ratio <= ratio <= max_ratio:
+            break
+
+        scale = math.sqrt(ratio / target_ratio)
+        if ratio < min_ratio:
+            scale = max(0.6, min(0.95, scale))
+        else:
+            scale = min(1.8, max(1.05, scale))
+
+        new_w = adjusted.w * scale
+        new_h = adjusted.h * scale
+        cx = adjusted.cx
+        cy = adjusted.cy
+        adjusted = Box(cx - new_w / 2.0, cy - new_h / 2.0, cx + new_w / 2.0, cy + new_h / 2.0)
+        adjusted = clamp_roi(adjusted, image_width, image_height)
+        if adjusted is None:
+            return None
+
+    return adjusted
+
+
+def boxes_in_roi(boxes: list[Box], roi: Box) -> list[Box]:
+    included: list[Box] = []
+    for box in boxes:
+        if not (roi.x1 <= box.cx <= roi.x2 and roi.y1 <= box.cy <= roi.y2):
+            continue
+        clipped = Box(
+            box.x1 - roi.x1,
+            box.y1 - roi.y1,
+            box.x2 - roi.x1,
+            box.y2 - roi.y1,
+        ).clip(roi.w, roi.h)
+        if clipped is not None and clipped.area > 4.0:
+            included.append(clipped)
+    return included
+
+
+def make_roi_samples(boxes: list[Box], image_width: int, image_height: int) -> list[RoiSample]:
+    samples: list[RoiSample] = []
+    groups = greedy_group_boxes(boxes)
+    for group in groups:
+        if len(group) == 2:
+            roi = expand_pair(boxes, group)
+            area_range = PAIR_AREA_RANGE
+            mode = "pair"
+        else:
+            roi = expand_single(boxes[group[0]])
+            area_range = SINGLE_AREA_RANGE
+            mode = "single"
+
+        roi = clamp_roi(roi, image_width, image_height)
+        if roi is None:
+            continue
+
+        object_area = sum(boxes[idx].area for idx in group)
+        roi = resize_roi_to_ratio(roi, image_width, image_height, object_area, *area_range)
+        if roi is None:
+            continue
+        samples.append(RoiSample(roi=roi, members=group, mode=mode))
+    return samples
+
+
+def to_yolo_lines(boxes: list[Box], roi_w: float, roi_h: float) -> list[str]:
+    lines: list[str] = []
+    for box in boxes:
+        xc = ((box.x1 + box.x2) / 2.0) / roi_w
+        yc = ((box.y1 + box.y2) / 2.0) / roi_h
+        bw = box.w / roi_w
+        bh = box.h / roi_h
+        lines.append(f"0 {xc:.6f} {yc:.6f} {bw:.6f} {bh:.6f}")
+    return lines
+
+
+def write_yaml(output_dir: Path, sources: list[str]) -> None:
+    yaml_path = output_dir / "data.yaml"
+    source_names = ", ".join(Path(item).name for item in sources)
+    yaml_path.write_text(
+        "\n".join(
+            [
+                "# ROI shoe training mix",
+                "",
+                f"path: {output_dir.resolve().as_posix()}",
+                "train: images/train",
+                "val: images/val",
+                "test: images/test",
+                "",
+                "nc: 1",
+                "names: ['shoe']",
+                "",
+                "dataset_info:",
+                "  name: shoe-roi-mix",
+                "  task: detect_shoe_roi",
+                f"  source: {source_names}",
+                "  note: cropped to foot-context ROIs to match online two-stage inference",
+                "",
+            ]
+        ),
+        encoding="utf-8",
+    )
+
+
+def build_split(source_dir: Path, output_dir: Path, split: str) -> dict[str, int]:
+    image_dir = source_dir / "images" / split
+    label_dir = source_dir / "labels" / split
+    if not image_dir.exists() or not label_dir.exists():
+        return {"images": 0, "boxes": 0, "single": 0, "pair": 0}
+
+    stats = {"images": 0, "boxes": 0, "single": 0, "pair": 0}
+    prefix = source_dir.name.replace("-", "_")
+
+    for label_path in sorted(label_dir.glob("*.txt")):
+        image_path = find_image(image_dir, label_path.stem)
+        if image_path is None:
+            continue
+
+        with Image.open(image_path) as image:
+            image = image.convert("RGB")
+            width, height = image.size
+            boxes = load_boxes(label_path, width, height)
+            if not boxes:
+                continue
+
+            samples = make_roi_samples(boxes, width, height)
+            for sample_idx, sample in enumerate(samples):
+                roi_boxes = boxes_in_roi(boxes, sample.roi)
+                if not roi_boxes:
+                    continue
+
+                out_stem = f"{prefix}_{label_path.stem}_{sample.mode}_{sample_idx:02d}"
+                dst_image = output_dir / "images" / split / f"{out_stem}.jpg"
+                dst_label = output_dir / "labels" / split / f"{out_stem}.txt"
+
+                crop = image.crop((sample.roi.x1, sample.roi.y1, sample.roi.x2, sample.roi.y2))
+                crop.save(dst_image, quality=95)
+
+                yolo_lines = to_yolo_lines(roi_boxes, sample.roi.w, sample.roi.h)
+                dst_label.write_text("\n".join(yolo_lines) + "\n", encoding="utf-8")
+
+                stats["images"] += 1
+                stats["boxes"] += len(roi_boxes)
+                stats[sample.mode] += 1
+
+    return stats
+
+
+def main() -> None:
+    args = parse_args()
+    output_dir = Path(args.output)
+
+    if args.clean and output_dir.exists():
+        shutil.rmtree(output_dir)
+
+    ensure_output_layout(output_dir)
+
+    summary: dict[str, dict[str, dict[str, int]]] = defaultdict(dict)
+    totals = {"images": 0, "boxes": 0, "single": 0, "pair": 0}
+
+    for source in args.sources:
+        source_dir = Path(source)
+        if not source_dir.exists():
+            raise FileNotFoundError(f"Source dataset not found: {source_dir}")
+
+        for split in ("train", "val", "test"):
+            stats = build_split(source_dir, output_dir, split)
+            summary[source_dir.name][split] = stats
+            for key in totals:
+                totals[key] += stats[key]
+
+    write_yaml(output_dir, args.sources)
+
+    print(f"Output dataset: {output_dir.resolve()}")
+    for source_name, split_map in summary.items():
+        print(f"[{source_name}]")
+        for split in ("train", "val", "test"):
+            stats = split_map.get(split, {"images": 0, "boxes": 0, "single": 0, "pair": 0})
+            print(
+                f"  {split}: rois={stats['images']} boxes={stats['boxes']} "
+                f"single={stats['single']} pair={stats['pair']}"
+            )
+
+    print(
+        "Total:"
+        f" rois={totals['images']} boxes={totals['boxes']}"
+        f" single={totals['single']} pair={totals['pair']}"
+    )
+
+
+if __name__ == "__main__":
+    main()

10_run_shoe_compare.ps1

@@ -1,6 +1,6 @@
 $ErrorActionPreference = "Stop"
 
-$repo = "C:\Users\tianj\Documents\apps\DetectionModelTraining"
+$repo = $PSScriptRoot
 $env:PYTHONPATH = "$repo\.pydeps"
 $env:YOLO_CONFIG_DIR = "$repo\.ultralytics"
 

11_run_shoe_compare_960.ps1

@@ -1,6 +1,6 @@
 $ErrorActionPreference = "Stop"
 
-$repo = "C:\Users\tianj\Documents\apps\DetectionModelTraining"
+$repo = $PSScriptRoot
 $env:PYTHONPATH = "$repo\.pydeps"
 $env:YOLO_CONFIG_DIR = "$repo\.ultralytics"
 

12_train_roi_yolov8s_640.bat

@@ -0,0 +1,92 @@
+@echo off
+chcp 65001 >nul
+cls
+
+set "REPO_DIR=%~dp0"
+pushd "%REPO_DIR%"
+
+set "YOLO_LAUNCHER="
+where yolo >nul 2>nul
+if %ERRORLEVEL% equ 0 set "YOLO_LAUNCHER=yolo"
+
+if not defined YOLO_LAUNCHER (
+    py -3.11 -c "import ultralytics" >nul 2>nul
+    if %ERRORLEVEL% equ 0 set "YOLO_LAUNCHER=py -3.11 -m ultralytics"
+)
+
+if not defined YOLO_LAUNCHER (
+    python -c "import ultralytics" >nul 2>nul
+    if %ERRORLEVEL% equ 0 set "YOLO_LAUNCHER=python -m ultralytics"
+)
+
+if not defined YOLO_LAUNCHER (
+    echo [错误] 未找到可用的 Ultralytics 启动方式
+    echo.
+    echo 请先确保满足以下任一条件:
+    echo   1. yolo 命令已加入 PATH
+    echo   2. py -3.11 可运行并已安装 ultralytics
+    echo   3. python 可运行并已安装 ultralytics
+    popd
+    pause
+    exit /b 1
+)
+
+set "DATASET=%REPO_DIR%datasets\shoe-roi-mix\data.yaml"
+
+if not exist "%DATASET%" (
+    echo [错误] 找不到 ROI 数据集配置: %DATASET%
+    echo.
+    echo 请先构建 ROI 数据集:
+    echo   python "%REPO_DIR%09_build_roi_shoe_dataset.py" --clean
+    popd
+    pause
+    exit /b 1
+)
+
+echo ============================================================
+echo      训练鞋子 ROI 检测模型 (YOLOv8s + 640x640)
+echo ============================================================
+echo.
+echo [信息] 数据集: %DATASET%
+echo [信息] 模型: yolov8s.pt
+echo.
+
+set "EPOCHS=150"
+set "IMGSZ=640"
+set "BATCH=16"
+set "PROJECT=%REPO_DIR%runs\roi_yolov8s_640"
+set "RUN_NAME=train_roi"
+
+echo 训练参数:
+echo   - Epochs: %EPOCHS%
+echo   - Image Size: %IMGSZ%x%IMGSZ%
+echo   - Batch Size: %BATCH%
+echo   - Device: GPU (cuda:0)
+echo   - Project: %PROJECT%
+echo   - Run Name: %RUN_NAME% ^(已存在时会自动递增，不覆盖旧模型^)
+echo.
+
+echo ============================================================
+echo                     开始训练
+echo ============================================================
+echo.
+
+call %YOLO_LAUNCHER% detect train data="%DATASET%" model="yolov8s.pt" epochs=%EPOCHS% imgsz=%IMGSZ% batch=%BATCH% device=0 project="%PROJECT%" name="%RUN_NAME%"
+
+if %ERRORLEVEL% neq 0 (
+    echo.
+    echo [错误] 训练失败！
+    popd
+    pause
+    exit /b 1
+)
+
+echo.
+echo ============================================================
+echo                     训练完成！
+echo ============================================================
+echo.
+echo 模型输出目录: %PROJECT%
+echo.
+popd
+pause

README.md

@@ -1,5 +1,68 @@
 # 鞋子检测模型训练指南
 
+## 当前主方案：YOLOv8s-640 + 脚部 ROI 训练
+
+当前项目的主训练方向已经调整为：
+- 只训练 `yolov8s`、输入尺寸固定 `640x640`
+- 训练数据不再直接使用“整张场景图”或“鞋子纯特写图”
+- 先根据鞋框裁出更接近线上输入分布的“脚部 ROI 图”，再训练鞋检测模型
+
+这样做的原因是线上链路并不是直接在整张图上找鞋，而是：
+1. 先从人体框生成脚部 ROI
+2. 再在脚部 ROI 上做鞋检测
+
+因此训练阶段也尽量模拟这个输入分布，保留一些裤脚、地面和周围背景，避免训练样本过于像商品特写。
+
+### ROI 规则
+
+单鞋 ROI：
+- 已知鞋框 `(x, y, w, h)`
+- `roi_x = x - 0.6w`
+- `roi_y = y - 0.5h`
+- `roi_w = 2.2w`
+- `roi_h = 2.4h`
+
+双鞋 ROI：
+- 优先把两只鞋裁进同一张 ROI
+- 先取两只鞋框并集，再扩框：
+- `roi_x = union_x - 0.35 * union_w`
+- `roi_y = union_y - 0.45 * union_h`
+- `roi_w = 1.7 * union_w`
+- `roi_h = 2.0 * union_h`
+
+裁图会自动裁剪到图像边界内。
+
+### 新主流程
+
+1. 准备原始单类鞋数据集
+
+```bash
+python 01_download_dataset.py --source openimages --max-samples 5000
+python 05_prepare_ppe_shoe_subset.py
+```
+
+2. 构建 ROI 化训练集
+
+```bash
+python 09_build_roi_shoe_dataset.py --clean
+```
+
+输出目录：
+- `datasets/shoe-roi-mix`
+
+3. 训练新的 ROI 模型
+
+```bash
+12_train_roi_yolov8s_640.bat
+```
+
+模型输出目录：
+- `runs/roi_yolov8s_640`
+
+说明：
+- 新模型会写到新的项目目录，不覆盖之前已有模型
+- 如果 `train_roi` 已存在，Ultralytics 会自动递增运行目录名
+
 ## 方案：640x640 单模型（部署时用2窗口）
 
 **训练阶段**：