项目初始化；已经训练了三种yolo版本的2种尺寸的模型

2026-03-16 09:55:52 +08:00 · 2026-03-16 09:55:52 +08:00 · 45f1f3182f
commit 45f1f3182f
14 changed files with 1588 additions and 0 deletions
--- a/.gitignore
+++ b/.gitignore
@ -0,0 +1,48 @@
+# Python cache and local tooling
+__pycache__/
+*.py[cod]
+*.pyo
+*.pyd
+.pytest_cache/
+.mypy_cache/
+.ruff_cache/
+.coverage
+.coverage.*
+
+# Virtualenvs and local dependency folders
+.venv/
+venv/
+env/
+ENV/
+.pydeps/
+
+# Project-local tool state
+.ultralytics/
+
+# OS / editor noise
+.DS_Store
+Thumbs.db
+Desktop.ini
+*.tmp
+*.temp
+
+# Large datasets and generated training assets
+datasets/
+runs/
+
+# Exported / downloaded model artifacts
+pretrained/
+*.pt
+*.pth
+*.onnx
+*.rknn
+*.engine
+*.bin
+
+# Logs and reports
+*.log
+
+# Keep placeholder files in otherwise ignored trees
+!datasets/.gitkeep
+!runs/.gitkeep
+!pretrained/.gitkeep
--- a/01_download_dataset.py
+++ b/01_download_dataset.py
@ -0,0 +1,411 @@
+#!/usr/bin/env python3
+"""
+下载鞋子检测数据集。
+
+支持:
+  - Ultralytics Construction-PPE
+  - Open Images V7 (推荐用于单类 shoe 检测)
+
+Open Images 推荐类别:
+  - Footwear
+  - Boot
+
+可选补充:
+  - Sandal
+
+不建议默认加入:
+  - High heels
+  - Roller skates
+"""
+
+import argparse
+import os
+import random
+import shutil
+import ssl
+import sys
+import urllib.request
+import zipfile
+from pathlib import Path
+
+
+OPENIMAGES_RECOMMENDED_CLASSES = ["Footwear", "Boot"]
+OPENIMAGES_OPTIONAL_CLASSES = ["Sandal"]
+OPENIMAGES_NOT_RECOMMENDED_CLASSES = ["High heels", "Roller skates"]
+
+
+def download_ultralytics_cppe(dataset_dir: str = "datasets/construction-ppe"):
+    """下载 Ultralytics Construction-PPE 数据集。"""
+    url = "https://github.com/ultralytics/assets/releases/download/v0.0.0/construction-ppe.zip"
+    zip_path = "construction-ppe.zip"
+
+    print("=" * 70)
+    print("下载 Construction-PPE 数据集")
+    print("=" * 70)
+    print(f"来源: {url}")
+    print(f"目标: {dataset_dir}")
+    print()
+
+    os.makedirs(dataset_dir, exist_ok=True)
+
+    print("[1/3] 下载中... (约 178MB)")
+    try:
+        ssl_context = ssl.create_default_context()
+        ssl_context.check_hostname = False
+        ssl_context.verify_mode = ssl.CERT_NONE
+
+        with urllib.request.urlopen(url, context=ssl_context, timeout=300) as response:
+            total_size = int(response.headers.get("content-length", 0))
+            downloaded = 0
+            chunk_size = 8192
+
+            with open(zip_path, "wb") as f:
+                while True:
+                    chunk = response.read(chunk_size)
+                    if not chunk:
+                        break
+                    f.write(chunk)
+                    downloaded += len(chunk)
+                    if total_size > 0:
+                        percent = (downloaded / total_size) * 100
+                        print(
+                            f"\r      进度: {percent:.1f}% ({downloaded}/{total_size} bytes)",
+                            end="",
+                        )
+
+        print("\n      ✓ 下载完成")
+    except Exception as e:
+        print(f"\n      ✗ 下载失败: {e}")
+        print("\n请手动下载:")
+        print(f"  1. 访问: {url}")
+        print("  2. 下载 construction-ppe.zip")
+        print(f"  3. 解压到 {dataset_dir}/")
+        return False
+
+    print("\n[2/3] 解压中...")
+    try:
+        with zipfile.ZipFile(zip_path, "r") as zip_ref:
+            zip_ref.extractall(dataset_dir)
+        print("      ✓ 解压完成")
+    except Exception as e:
+        print(f"      ✗ 解压失败: {e}")
+        return False
+
+    print("\n[3/3] 清理临时文件...")
+    os.remove(zip_path)
+    print("      ✓ 完成")
+    return True
+
+
+def create_single_class_yaml(dataset_dir: str, source_name: str, dataset_path_value: str = "."):
+    """创建单类 shoe 检测配置文件。"""
+    yaml_content = f"""# 单类鞋子检测数据集配置
+
+path: {dataset_path_value}
+train: images/train
+val: images/val
+test: images/test
+
+nc: 1
+names: ['shoe']
+
+dataset_info:
+  name: {source_name}
+  task: detect_shoe
+  note: 所有鞋类子类统一映射为单一类别 shoe
+"""
+
+    yaml_path = os.path.join(dataset_dir, "data.yaml")
+    with open(yaml_path, "w", encoding="utf-8") as f:
+        f.write(yaml_content)
+
+    print(f"\n✓ 配置文件创建: {yaml_path}")
+    return yaml_path
+
+
+def rewrite_yaml_for_existing_splits(dataset_dir: str, source_name: str):
+    """根据现有目录结构重写 data.yaml。"""
+    images_root = Path(dataset_dir) / "images"
+    split_names = [name for name in ("train", "val", "test") if (images_root / name).exists()]
+
+    if not split_names:
+        raise RuntimeError(f"未找到任何图像 split: {images_root}")
+
+    train_split = "train" if "train" in split_names else split_names[0]
+    val_split = "val" if "val" in split_names else train_split
+    test_line = f"test: images/{'test' if 'test' in split_names else val_split}"
+
+    yaml_content = f"""# 单类鞋子检测数据集配置
+
+path: .
+train: images/{train_split}
+val: images/{val_split}
+{test_line}
+
+nc: 1
+names: ['shoe']
+
+dataset_info:
+  name: {source_name}
+  task: detect_shoe
+  note: 所有鞋类子类统一映射为单一类别 shoe
+"""
+
+    yaml_path = Path(dataset_dir) / "data.yaml"
+    yaml_path.write_text(yaml_content, encoding="utf-8")
+    print(f"\n✓ 配置文件更新: {yaml_path}")
+    return str(yaml_path)
+
+
+def ensure_openimages_train_val_split(export_dir: str, train_ratio: float = 0.9, seed: int = 42):
+    """如果导出结果只有单个 split，则自动切分为 train/val。"""
+    images_root = Path(export_dir) / "images"
+    labels_root = Path(export_dir) / "labels"
+    train_images = images_root / "train"
+    val_images = images_root / "val"
+    train_labels = labels_root / "train"
+    val_labels = labels_root / "val"
+
+    if train_images.exists() and train_labels.exists():
+        return
+
+    if not val_images.exists() or not val_labels.exists():
+        return
+
+    image_files = sorted([p for p in val_images.iterdir() if p.is_file()])
+    if len(image_files) < 2:
+        return
+
+    train_images.mkdir(parents=True, exist_ok=True)
+    train_labels.mkdir(parents=True, exist_ok=True)
+
+    rng = random.Random(seed)
+    rng.shuffle(image_files)
+
+    split_idx = max(1, min(len(image_files) - 1, int(len(image_files) * train_ratio)))
+    train_files = image_files[:split_idx]
+
+    for image_path in train_files:
+        label_path = val_labels / f"{image_path.stem}.txt"
+        shutil.move(str(image_path), train_images / image_path.name)
+        if label_path.exists():
+            shutil.move(str(label_path), train_labels / label_path.name)
+
+    print("\n自动切分单一 split 为 train/val:")
+    print(f"  train: {len(list(train_images.iterdir()))} 张")
+    print(f"  val: {len(list(val_images.iterdir()))} 张")
+
+
+def merge_openimages_to_single_class(export_dir: str):
+    """将 Open Images 导出的多类别 YOLO 标签合并为单类 shoe。"""
+    labels_root = Path(export_dir) / "labels"
+    if not labels_root.exists():
+        print(f"✗ 未找到标签目录: {labels_root}")
+        return False
+
+    total_files = 0
+    total_boxes = 0
+    changed_files = 0
+
+    for label_file in labels_root.rglob("*.txt"):
+        total_files += 1
+        lines = label_file.read_text(encoding="utf-8").splitlines()
+        rewritten = []
+        file_changed = False
+
+        for line in lines:
+            parts = line.strip().split()
+            if len(parts) < 5:
+                continue
+            parts[0] = "0"
+            rewritten.append(" ".join(parts))
+            total_boxes += 1
+            file_changed = True
+
+        label_file.write_text("\n".join(rewritten) + ("\n" if rewritten else ""), encoding="utf-8")
+
+        if file_changed:
+            changed_files += 1
+
+    ensure_openimages_train_val_split(export_dir)
+    rewrite_yaml_for_existing_splits(export_dir, "Open Images V7")
+
+    print("\n单类合并完成:")
+    print(f"  标签文件: {changed_files}/{total_files}")
+    print(f"  标注框数: {total_boxes}")
+    return True
+
+
+def download_openimages(classes: list, max_samples: int, dataset_dir: str):
+    """通过 FiftyOne 下载 Open Images 并导出为单类 shoe 数据集。"""
+    try:
+        import fiftyone as fo
+        import fiftyone.zoo as foz
+    except ImportError:
+        print("错误: 未安装 fiftyone")
+        print("请运行: pip install fiftyone fiftyone-db-ubuntu2204")
+        return False
+
+    export_dir = dataset_dir + "-yolo"
+
+    print("=" * 70)
+    print("下载 Open Images V7 数据集")
+    print("=" * 70)
+    print(f"类别: {classes}")
+    print(f"最大样本数: {max_samples}")
+    print("原始缓存目录: FiftyOne 默认缓存目录")
+    print(f"YOLO 导出目录: {export_dir}")
+    print()
+
+    if os.path.exists(export_dir):
+        print(f"检测到已有导出目录，先删除: {export_dir}")
+        shutil.rmtree(export_dir)
+
+    try:
+        dataset = foz.load_zoo_dataset(
+            "open-images-v7",
+            split="train",
+            label_types=["detections"],
+            classes=classes,
+            max_samples=max_samples,
+        )
+
+        print("\n导出为 YOLO 格式...")
+        dataset.export(
+            export_dir=export_dir,
+            dataset_type=fo.types.YOLOv5Dataset,
+            label_field="ground_truth",
+        )
+
+        if not merge_openimages_to_single_class(export_dir):
+            return False
+
+        print(f"\n✓ 数据集保存: {export_dir}")
+        return True
+    except Exception as e:
+        print(f"✗ 下载失败: {e}")
+        return False
+
+
+def check_dataset(dataset_dir: str):
+    """检查数据集完整性。"""
+    print("\n" + "=" * 70)
+    print("检查数据集")
+    print("=" * 70)
+
+    required_dirs = [
+        "images/train",
+        "images/val",
+        "images/test",
+        "labels/train",
+        "labels/val",
+        "labels/test",
+    ]
+
+    all_ok = True
+    for dir_name in required_dirs:
+        full_path = os.path.join(dataset_dir, dir_name)
+        if os.path.exists(full_path):
+            count = len(
+                [f for f in os.listdir(full_path) if os.path.isfile(os.path.join(full_path, f))]
+            )
+            print(f"  ✓ {dir_name}: {count} 个文件")
+        else:
+            print(f"  ✗ {dir_name}: 不存在")
+            all_ok = False
+
+    yaml_path = os.path.join(dataset_dir, "data.yaml")
+    if os.path.exists(yaml_path):
+        print(f"  ✓ data.yaml: 已生成")
+    else:
+        print(f"  ✗ data.yaml: 不存在")
+        all_ok = False
+
+    return all_ok
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="下载鞋子检测数据集",
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+        epilog="""
+示例:
+  # 下载 Construction-PPE
+  python 01_download_dataset.py --source ultralytics
+
+  # 下载 Open Images 推荐鞋子类别
+  python 01_download_dataset.py --source openimages --max-samples 5000
+
+  # 如需补充凉鞋样本
+  python 01_download_dataset.py --source openimages --classes Footwear Boot Sandal --max-samples 8000
+        """,
+    )
+
+    parser.add_argument(
+        "--source",
+        choices=["ultralytics", "openimages"],
+        default="openimages",
+        help="数据源 (默认: openimages)",
+    )
+    parser.add_argument(
+        "--dir",
+        default="datasets/openimages-shoes",
+        help="数据集保存目录",
+    )
+    parser.add_argument(
+        "--max-samples",
+        type=int,
+        default=5000,
+        help="Open Images 最大样本数 (默认: 5000)",
+    )
+    parser.add_argument(
+        "--classes",
+        nargs="+",
+        default=OPENIMAGES_RECOMMENDED_CLASSES,
+        help="Open Images 类别 (默认: Footwear Boot)",
+    )
+
+    args = parser.parse_args()
+    success = False
+    final_dataset_dir = args.dir
+
+    print("=" * 70)
+    print("鞋子检测数据集准备")
+    print("=" * 70)
+    if args.source == "openimages":
+        print(f"推荐类别: {OPENIMAGES_RECOMMENDED_CLASSES}")
+        print(f"可选补充: {OPENIMAGES_OPTIONAL_CLASSES}")
+        print(f"默认不建议: {OPENIMAGES_NOT_RECOMMENDED_CLASSES}")
+        print()
+
+    if args.source == "ultralytics":
+        success = download_ultralytics_cppe(args.dir)
+        if success:
+            create_single_class_yaml(args.dir, "Construction-PPE", dataset_path_value="construction-ppe")
+            check_dataset(args.dir)
+
+    elif args.source == "openimages":
+        success = download_openimages(args.classes, args.max_samples, args.dir)
+        final_dataset_dir = args.dir + "-yolo"
+        if success:
+            check_dataset(final_dataset_dir)
+
+    if success:
+        print("\n" + "=" * 70)
+        print("数据集准备完成！")
+        print("=" * 70)
+        print(f"训练数据集路径: {final_dataset_dir}")
+        print("\n下一步:")
+        print(f"  1. 检查配置: {final_dataset_dir}/data.yaml")
+        print(f"  2. 开始训练: 02_train.bat")
+        print(
+            f"  3. 或手动: yolo detect train data={final_dataset_dir}/data.yaml model=yolov8s.pt epochs=150 imgsz=640"
+        )
+        return 0
+
+    print("\n✗ 数据集准备失败")
+    return 1
+
+
+if __name__ == "__main__":
+    sys.exit(main())
--- a/02_train.bat
+++ b/02_train.bat
@ -0,0 +1,90 @@
+@echo off
+chcp 65001 >nul
+cls
+
+:: 设置 Python 3.11 路径
+set "PATH=C:\Users\Tellme\AppData\Local\Programs\Python\Python311\Scripts;C:\Users\Tellme\AppData\Local\Programs\Python\Python311;%PATH%"
+
+echo ============================================================
+echo           训练鞋子检测模型 (YOLOv8 + 640x640)
+echo ============================================================
+echo.
+
+:: 设置数据集路径
+set DATASET=datasets/openimages-shoes-yolo/data.yaml
+
+:: 检查数据集是否存在
+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
+    pause
+    exit /b 1
+)
+
+echo [信息] 数据集: %DATASET%
+echo.
+
+:: 选择模型
+echo 选择模型:
+echo   1. YOLOv8n (轻量级, 速度快)
+echo   2. YOLOv8s (推荐, 速度和精度平衡)
+echo   3. YOLOv8m (高精度, 较慢)
+echo.
+set /p MODEL_CHOICE="输入选择 (1-3, 默认 2): "
+
+if "%MODEL_CHOICE%"=="" set MODEL_CHOICE=2
+if "%MODEL_CHOICE%"=="1" (
+    set MODEL=yolov8n.pt
+    set DESC=YOLOv8n
+)
+if "%MODEL_CHOICE%"=="2" (
+    set MODEL=yolov8s.pt
+    set DESC=YOLOv8s (推荐)
+)
+if "%MODEL_CHOICE%"=="3" (
+    set MODEL=yolov8m.pt
+    set DESC=YOLOv8m
+)
+
+echo.
+echo [信息] 使用模型: %DESC%
+echo.
+
+:: 训练参数
+set EPOCHS=150
+set IMGSZ=640
+set BATCH=16
+
+echo 训练参数:
+echo   - Epochs: %EPOCHS%
+echo   - Image Size: %IMGSZ%x%IMGSZ%
+echo   - Batch Size: %BATCH%
+echo   - Device: GPU (cuda:0)
+echo.
+
+echo ============================================================
+echo                     开始训练
+echo ============================================================
+echo.
+
+yolo 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！
+    pause
+    exit /b 1
+)
+
+echo.
+echo ============================================================
+echo                     训练完成！
+echo ============================================================
+echo.
+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.
+pause
--- a/03_export_onnx.bat
+++ b/03_export_onnx.bat
@ -0,0 +1,38 @@
+@echo off
+chcp 65001 >nul
+cls
+
+:: 设置 Python 3.11 路径
+set "PATH=C:\Users\Tellme\AppData\Local\Programs\Python\Python311\Scripts;C:\Users\Tellme\AppData\Local\Programs\Python\Python311;%PATH%"
+
+echo ============================================================
+echo         导出 ONNX 模型 (640x640)
+echo ============================================================
+echo.
+
+set MODEL_PATH=%USERPROFILE%\apps\ultralytics\runs\detect\train\weights\best.pt
+
+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
+    pause
+    exit /b 1
+)
+
+echo [信息] 输入模型: %MODEL_PATH%
+echo.
+
+yolo export model=%MODEL_PATH% format=onnx imgsz=640 opset=12 simplify
+
+if %ERRORLEVEL% neq 0 (
+    echo [错误] 导出失败！
+    pause
+    exit /b 1
+)
+
+echo.
+echo [成功] ONNX 模型: %USERPROFILE%\apps\ultralytics\runs\detect\train\weights\best.onnx
+echo.
+echo 下一步: 在 Ubuntu 上运行 04_convert_rknn.py 转换
+echo.
+pause
--- a/04_convert_rknn.py
+++ b/04_convert_rknn.py
@ -0,0 +1,262 @@
+#!/usr/bin/env python3
+"""
+将 YOLOv8 ONNX 模型转换为 RKNN 格式
+适用于 RK3588 / RK3568 / RK3576 等平台
+
+环境要求:
+    - Ubuntu x86_64 / Docker
+    - Python 3.8 / 3.9 / 3.10 / 3.11
+    - rknn-toolkit2 (pip install rknn-toolkit2==2.2.0)
+
+使用方法:
+    # FP16 模式（推荐，速度快精度高）
+    python 04_convert_rknn.py best.onnx -o shoe_detector.rknn -t rk3588
+    
+    # INT8 量化（模型更小，需要校准数据集）
+    python 04_convert_rknn.py best.onnx -o shoe_detector.rknn -t rk3588 -q -d dataset.txt
+
+支持的 target_platform:
+    - rk3588 / rk3588s
+    - rk3568 / rk3566
+    - rk3576
+    - rv1106 / rv1103 / rv1103b
+    - rv1126
+"""
+
+import argparse
+import os
+import sys
+from pathlib import Path
+
+
+def check_environment():
+    """检查运行环境"""
+    try:
+        from rknn.api import RKNN
+        print("✓ RKNN Toolkit2 已安装")
+        return True
+    except ImportError:
+        print("✗ 错误: 未安装 RKNN Toolkit2")
+        print("\n请安装:")
+        print("  pip install rknn-toolkit2==2.2.0")
+        print("\n或从源码安装:")
+        print("  https://github.com/airockchip/rknn-toolkit2")
+        return False
+
+
+def create_sample_dataset(onnx_path: str, output_path: str = "dataset.txt", num_samples: int = 20):
+    """
+    创建示例量化校准数据集
+    用于 INT8 量化时提供校准图片路径
+    """
+    print(f"\n创建示例校准数据集: {output_path}")
+    print("注意: 请用实际图片替换这些示例路径")
+    
+    sample_content = f"""# RKNN INT8 量化校准数据集
+# 每行一个图片路径，建议使用 20-100 张典型场景图片
+# 图片格式: JPG, PNG, BMP 等
+
+# 示例路径（请替换为实际路径）:
+# /path/to/train/images/img001.jpg
+# /path/to/train/images/img002.jpg
+# /path/to/valid/images/img001.jpg
+
+# 提示:
+# 1. 图片应与实际部署场景相似
+# 2. 包含各种光照、角度、背景的样本
+# 3. 建议 20-100 张，越多越慢但可能更准
+"""
+    
+    with open(output_path, 'w') as f:
+        f.write(sample_content)
+    
+    print(f"✓ 示例数据集已创建: {output_path}")
+    print("  请编辑此文件，添加实际的图片路径")
+    return output_path
+
+
+def convert_onnx_to_rknn(
+    onnx_path: str,
+    output_path: str = None,
+    target_platform: str = "rk3588",
+    do_quantization: bool = False,
+    dataset_path: str = None,
+    verbose: bool = True
+):
+    """
+    转换 ONNX 模型到 RKNN
+    
+    Args:
+        onnx_path: ONNX 模型文件路径
+        output_path: 输出 RKNN 文件路径，默认与 ONNX 同名
+        target_platform: 目标平台，默认 rk3588
+        do_quantization: 是否启用 INT8 量化
+        dataset_path: 量化校准数据集路径（txt 文件，每行一张图片路径）
+        verbose: 是否打印详细信息
+    """
+    if output_path is None:
+        output_path = onnx_path.replace(".onnx", ".rknn")
+    
+    # 确保输出目录存在
+    output_dir = os.path.dirname(output_path)
+    if output_dir and not os.path.exists(output_dir):
+        os.makedirs(output_dir)
+    
+    print("="*70)
+    print(f"ONNX 转 RKNN")
+    print("="*70)
+    print(f"输入:  {onnx_path}")
+    print(f"输出:  {output_path}")
+    print(f"目标:  {target_platform}")
+    print(f"量化:  {'INT8' if do_quantization else 'FP16 (无量化)'}")
+    if do_quantization:
+        print(f"校准:  {dataset_path}")
+    print("="*70)
+    
+    # 检查输入文件
+    if not os.path.exists(onnx_path):
+        print(f"\n✗ 错误: 找不到 ONNX 文件: {onnx_path}")
+        return False
+    
+    # 检查数据集（如果需要量化）
+    if do_quantization:
+        if dataset_path is None:
+            print("\n✗ 错误: INT8 量化需要提供校准数据集")
+            print("  使用 --dataset 指定数据集文件路径")
+            print("  或运行 --create-dataset 创建示例")
+            return False
+        if not os.path.exists(dataset_path):
+            print(f"\n✗ 错误: 找不到数据集文件: {dataset_path}")
+            return False
+    
+    from rknn.api import RKNN
+    
+    # 创建 RKNN 对象
+    rknn = RKNN(verbose=verbose)
+    
+    try:
+        # 配置模型
+        print("\n[1/4] 配置模型...")
+        rknn.config(
+            mean_values=[[0, 0, 0]],           # YOLOv8 使用 0-255 输入
+            std_values=[[255, 255, 255]],      # 归一化到 0-1
+            target_platform=target_platform
+        )
+        print("      ✓ 完成")
+        
+        # 加载 ONNX
+        print("\n[2/4] 加载 ONNX 模型...")
+        ret = rknn.load_onnx(model=onnx_path)
+        if ret != 0:
+            print("      ✗ 加载失败!")
+            return False
+        print("      ✓ 完成")
+        
+        # 构建模型
+        print("\n[3/4] 构建 RKNN 模型...")
+        if do_quantization:
+            print(f"      使用 INT8 量化，校准数据集: {dataset_path}")
+            ret = rknn.build(do_quantization=True, dataset=dataset_path)
+        else:
+            print("      使用 FP16 模式（无量化）")
+            ret = rknn.build(do_quantization=False)
+        
+        if ret != 0:
+            print("      ✗ 构建失败!")
+            return False
+        print("      ✓ 完成")
+        
+        # 导出 RKNN
+        print("\n[4/4] 导出 RKNN 模型...")
+        ret = rknn.export_rknn(output_path)
+        if ret != 0:
+            print("      ✗ 导出失败!")
+            return False
+        print("      ✓ 完成")
+        
+    finally:
+        rknn.release()
+    
+    # 验证输出
+    if os.path.exists(output_path):
+        size_mb = os.path.getsize(output_path) / (1024 * 1024)
+        print("\n" + "="*70)
+        print(f"✓ 转换成功!")
+        print(f"  输出文件: {output_path}")
+        print(f"  文件大小: {size_mb:.2f} MB")
+        print("="*70)
+        
+        print("\n下一步:")
+        print(f"  1. 复制到 RK3588:")
+        print(f"     scp {output_path} orangepi@<rk3588_ip>:/home/orangepi/apps/OrangePi3588Media/models/")
+        print(f"  2. 更新配置文件中的模型路径")
+        return True
+    else:
+        print("\n✗ 错误: 输出文件未生成")
+        return False
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="将 YOLOv8 ONNX 模型转换为 RKNN",
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+        epilog="""
+示例:
+  # FP16 模式（推荐）
+  python 04_convert_rknn.py best.onnx -o shoe_detector.rknn
+  
+  # 指定目标平台
+  python 04_convert_rknn.py best.onnx -t rk3568
+  
+  # INT8 量化
+  python 04_convert_rknn.py best.onnx -q -d dataset.txt
+  
+  # 创建示例校准数据集
+  python 04_convert_rknn.py --create-dataset
+        """
+    )
+    
+    parser.add_argument("onnx", nargs="?", help="ONNX 模型文件路径")
+    parser.add_argument("-o", "--output", help="输出 RKNN 文件路径")
+    parser.add_argument("-t", "--target", default="rk3588",
+                       help="目标平台 (默认: rk3588)")
+    parser.add_argument("-q", "--quantize", action="store_true",
+                       help="启用 INT8 量化")
+    parser.add_argument("-d", "--dataset", help="量化校准数据集路径 (txt 文件)")
+    parser.add_argument("--create-dataset", action="store_true",
+                       help="创建示例校准数据集并退出")
+    parser.add_argument("-v", "--verbose", action="store_true", default=True,
+                       help="显示详细信息")
+    
+    args = parser.parse_args()
+    
+    # 创建示例数据集
+    if args.create_dataset:
+        create_sample_dataset("best.onnx", "dataset.txt")
+        return 0
+    
+    # 检查参数
+    if args.onnx is None:
+        parser.print_help()
+        print("\n错误: 请提供 ONNX 文件路径")
+        return 1
+    
+    # 检查环境
+    if not check_environment():
+        return 1
+    
+    # 执行转换
+    success = convert_onnx_to_rknn(
+        onnx_path=args.onnx,
+        output_path=args.output,
+        target_platform=args.target,
+        do_quantization=args.quantize,
+        dataset_path=args.dataset,
+        verbose=args.verbose
+    )
+    
+    return 0 if success else 1
+
+
+if __name__ == "__main__":
+    sys.exit(main())
--- a/05_prepare_ppe_shoe_subset.py
+++ b/05_prepare_ppe_shoe_subset.py
@ -0,0 +1,119 @@
+#!/usr/bin/env python3
+"""从 Construction-PPE 中提取单类 shoe 子集。"""
+
+import argparse
+import shutil
+from pathlib import Path
+
+
+SHOE_CLASSES = {"3", "10"}  # boots, no_boots
+
+
+def ensure_clean_dir(path: Path):
+    if path.exists():
+        shutil.rmtree(path)
+    path.mkdir(parents=True, exist_ok=True)
+
+
+def write_yaml(output_dir: Path):
+    yaml_path = output_dir / "data.yaml"
+    abs_output = output_dir.resolve().as_posix()
+    yaml_path.write_text(
+        "\n".join(
+            [
+                "# PPE 鞋子单类微调数据集配置",
+                "",
+                f"path: {abs_output}",
+                "train: images/train",
+                "val: images/val",
+                "test: images/test",
+                "",
+                "nc: 1",
+                "names: ['shoe']",
+                "",
+                "dataset_info:",
+                "  name: Construction-PPE shoe subset",
+                "  source: Construction-PPE",
+                "  note: 从 boots 和 no_boots 提取并统一映射为 shoe",
+                "",
+            ]
+        ),
+        encoding="utf-8",
+    )
+
+
+def convert_split(source_dir: Path, output_dir: Path, split: str):
+    image_src = source_dir / "images" / split
+    label_src = source_dir / "labels" / split
+    image_dst = output_dir / "images" / split
+    label_dst = output_dir / "labels" / split
+
+    image_dst.mkdir(parents=True, exist_ok=True)
+    label_dst.mkdir(parents=True, exist_ok=True)
+
+    kept_images = 0
+    kept_boxes = 0
+
+    for label_file in sorted(label_src.glob("*.txt")):
+        lines = label_file.read_text(encoding="utf-8").splitlines()
+        shoe_lines = []
+
+        for line in lines:
+            parts = line.strip().split()
+            if len(parts) < 5 or parts[0] not in SHOE_CLASSES:
+                continue
+            parts[0] = "0"
+            shoe_lines.append(" ".join(parts))
+
+        if not shoe_lines:
+            continue
+
+        image_file = image_src / f"{label_file.stem}.jpg"
+        if not image_file.exists():
+            image_file = image_src / f"{label_file.stem}.png"
+        if not image_file.exists():
+            continue
+
+        shutil.copy2(image_file, image_dst / image_file.name)
+        (label_dst / label_file.name).write_text("\n".join(shoe_lines) + "\n", encoding="utf-8")
+        kept_images += 1
+        kept_boxes += len(shoe_lines)
+
+    return kept_images, kept_boxes
+
+
+def main():
+    parser = argparse.ArgumentParser(description="提取 PPE 鞋子单类子集")
+    parser.add_argument(
+        "--source",
+        default="datasets/construction-ppe",
+        help="Construction-PPE 数据集目录",
+    )
+    parser.add_argument(
+        "--output",
+        default="datasets/ppe-shoes",
+        help="输出目录",
+    )
+    args = parser.parse_args()
+
+    source_dir = Path(args.source)
+    output_dir = Path(args.output)
+
+    ensure_clean_dir(output_dir)
+
+    total_images = 0
+    total_boxes = 0
+    for split in ("train", "val", "test"):
+        kept_images, kept_boxes = convert_split(source_dir, output_dir, split)
+        total_images += kept_images
+        total_boxes += kept_boxes
+        print(f"[{split}] images={kept_images} boxes={kept_boxes}")
+
+    write_yaml(output_dir)
+    print(f"\n输出目录: {output_dir}")
+    print(f"总图片数: {total_images}")
+    print(f"总框数: {total_boxes}")
+
+
+if __name__ == "__main__":
+    main()
--- a/06_finetune_ppe.bat
+++ b/06_finetune_ppe.bat
@ -0,0 +1,62 @@
+@echo off
+setlocal
+
+set "PATH=C:\Users\Tellme\AppData\Local\Programs\Python\Python311\Scripts;C:\Users\Tellme\AppData\Local\Programs\Python\Python311;%PATH%"
+
+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 "BASE_MODEL=C:\Users\Tellme\apps\ultralytics\runs\detect\train3\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
+    pause
+    exit /b 1
+)
+
+if not exist "%BASE_MODEL%" (
+    echo [ERROR] Base model not found: %BASE_MODEL%
+    pause
+    exit /b 1
+)
+
+echo [INFO] Dataset: %DATASET%
+echo [INFO] Base model: %BASE_MODEL%
+echo.
+
+set "EPOCHS=50"
+set "IMGSZ=640"
+set "BATCH=16"
+
+echo Fine-tune params:
+echo   Epochs: %EPOCHS%
+echo   Image Size: %IMGSZ%x%IMGSZ%
+echo   Batch Size: %BATCH%
+echo   Device: GPU (cuda:0)
+echo.
+
+echo ============================================================
+echo                    Start fine-tuning
+echo ============================================================
+echo.
+
+yolo 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.
+    pause
+    exit /b 1
+)
+
+echo.
+echo ============================================================
+echo                    Fine-tuning complete
+echo ============================================================
+echo.
+pause
--- a/07_build_public_shoe_dataset.py
+++ b/07_build_public_shoe_dataset.py
@ -0,0 +1,145 @@
+#!/usr/bin/env python3
+"""Build a merged single-class public shoe dataset for YOLO training."""
+
+from __future__ import annotations
+
+import argparse
+import shutil
+from collections import defaultdict
+from pathlib import Path
+
+
+DEFAULT_SOURCES = [
+    "datasets/openimages-shoes-yolo",
+    "datasets/ppe-shoes",
+]
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Merge public shoe datasets into one YOLO dataset")
+    parser.add_argument(
+        "--sources",
+        nargs="+",
+        default=DEFAULT_SOURCES,
+        help="Source dataset directories containing images/<split> and labels/<split>",
+    )
+    parser.add_argument(
+        "--output",
+        default="datasets/shoe-public-mix",
+        help="Output merged 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 copy_split(source_dir: Path, output_dir: Path, split: str) -> tuple[int, int]:
+    image_dir = source_dir / "images" / split
+    label_dir = source_dir / "labels" / split
+    if not image_dir.exists() or not label_dir.exists():
+        return 0, 0
+
+    images_copied = 0
+    boxes_copied = 0
+    prefix = source_dir.name.replace("-", "_")
+
+    for label_file in sorted(label_dir.glob("*.txt")):
+        image_file = None
+        for ext in (".jpg", ".jpeg", ".png", ".bmp", ".webp"):
+            candidate = image_dir / f"{label_file.stem}{ext}"
+            if candidate.exists():
+                image_file = candidate
+                break
+
+        if image_file is None:
+            continue
+
+        lines = [line.strip() for line in label_file.read_text(encoding="utf-8").splitlines() if line.strip()]
+        if not lines:
+            continue
+
+        out_stem = f"{prefix}_{label_file.stem}"
+        dst_image = output_dir / "images" / split / f"{out_stem}{image_file.suffix.lower()}"
+        dst_label = output_dir / "labels" / split / f"{out_stem}.txt"
+
+        shutil.copy2(image_file, dst_image)
+        dst_label.write_text("\n".join(lines) + "\n", encoding="utf-8")
+
+        images_copied += 1
+        boxes_copied += len(lines)
+
+    return images_copied, boxes_copied
+
+
+def write_yaml(output_dir: Path) -> None:
+    yaml_path = output_dir / "data.yaml"
+    yaml_path.write_text(
+        "\n".join(
+            [
+                "# Public 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-public-mix",
+                "  task: detect_shoe",
+                "  note: merged Open Images shoe data and PPE shoe subset",
+                "",
+            ]
+        ),
+        encoding="utf-8",
+    )
+
+
+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, tuple[int, int]]] = defaultdict(dict)
+    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"):
+            summary[source_dir.name][split] = copy_split(source_dir, output_dir, split)
+
+    write_yaml(output_dir)
+
+    print(f"Output dataset: {output_dir.resolve()}")
+    total_images = 0
+    total_boxes = 0
+    for source_name, split_map in summary.items():
+        print(f"[{source_name}]")
+        for split in ("train", "val", "test"):
+            images_copied, boxes_copied = split_map.get(split, (0, 0))
+            total_images += images_copied
+            total_boxes += boxes_copied
+            print(f"  {split}: images={images_copied} boxes={boxes_copied}")
+
+    print(f"Total images: {total_images}")
+    print(f"Total boxes: {total_boxes}")
+
+
+if __name__ == "__main__":
+    main()
--- a/08_train_compare_models.py
+++ b/08_train_compare_models.py
@ -0,0 +1,155 @@
+#!/usr/bin/env python3
+"""Train multiple YOLO models on the merged shoe dataset and compare ROI performance."""
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import sys
+from pathlib import Path
+from statistics import mean
+
+
+REPO_ROOT = Path(__file__).resolve().parent
+PYDEPS = REPO_ROOT / ".pydeps"
+if PYDEPS.exists():
+    sys.path.insert(0, str(PYDEPS))
+
+from ultralytics import YOLO  # noqa: E402
+
+
+DEFAULT_MODELS = ["yolov8s.pt", "yolo11s.pt", "yolo26s.pt"]
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Train and compare YOLO shoe detectors")
+    parser.add_argument("--data", default="datasets/shoe-public-mix/data.yaml", help="Training dataset yaml")
+    parser.add_argument("--roi-dir", default="datasets/roi-shoes", help="Real-world ROI evaluation directory")
+    parser.add_argument("--project", default="runs/shoe_compare", help="Ultralytics project directory")
+    parser.add_argument("--models", nargs="+", default=DEFAULT_MODELS, help="Model checkpoints or aliases")
+    parser.add_argument("--epochs", type=int, default=40, help="Training epochs for each model")
+    parser.add_argument("--imgsz", type=int, default=640, help="Training image size")
+    parser.add_argument("--batch", type=int, default=16, help="Training batch size")
+    parser.add_argument("--workers", type=int, default=8, help="Data loader workers")
+    parser.add_argument("--device", default="0", help="Training device, e.g. 0 or cpu")
+    parser.add_argument("--seed", type=int, default=42, help="Random seed")
+    parser.add_argument("--patience", type=int, default=20, help="Early stopping patience")
+    return parser.parse_args()
+
+
+def safe_name(model_name: str) -> str:
+    stem = Path(model_name).stem
+    return stem.replace(".", "_").replace("-", "_")
+
+
+def evaluate_roi(model_path: Path, roi_dir: Path, save_dir: Path, device: str) -> dict:
+    model = YOLO(str(model_path))
+    results = model.predict(
+        source=str(roi_dir),
+        conf=0.1,
+        save=True,
+        project=str(save_dir.parent),
+        name=save_dir.name,
+        exist_ok=True,
+        verbose=False,
+        device=device,
+    )
+
+    per_image = []
+    max_confs = []
+    hit_count = 0
+    for result in results:
+        n = 0 if result.boxes is None else len(result.boxes)
+        hit = n > 0
+        if hit:
+            hit_count += 1
+            max_confs.append(float(result.boxes.conf.max().item()))
+        per_image.append(
+            {
+                "image": Path(result.path).name,
+                "detections": n,
+                "max_conf": float(result.boxes.conf.max().item()) if hit else 0.0,
+            }
+        )
+
+    summary = {
+        "roi_total": len(per_image),
+        "roi_hits": hit_count,
+        "roi_hit_rate": hit_count / len(per_image) if per_image else 0.0,
+        "roi_mean_max_conf": mean(max_confs) if max_confs else 0.0,
+        "per_image": per_image,
+    }
+    return summary
+
+
+def main() -> None:
+    args = parse_args()
+    os.environ.setdefault("YOLO_CONFIG_DIR", str(REPO_ROOT / ".ultralytics"))
+
+    data = Path(args.data)
+    roi_dir = Path(args.roi_dir)
+    project = Path(args.project)
+    project.mkdir(parents=True, exist_ok=True)
+
+    if not data.exists():
+        raise FileNotFoundError(f"Dataset yaml not found: {data}")
+    if not roi_dir.exists():
+        raise FileNotFoundError(f"ROI directory not found: {roi_dir}")
+
+    compare_summary = []
+    for model_name in args.models:
+        run_name = f"{safe_name(model_name)}_shoe_{args.imgsz}"
+        print("=" * 80)
+        print(f"Training {model_name} -> {run_name}")
+        print("=" * 80)
+
+        model = YOLO(model_name)
+        train_results = model.train(
+            data=str(data.resolve()),
+            epochs=args.epochs,
+            imgsz=args.imgsz,
+            batch=args.batch,
+            workers=args.workers,
+            device=args.device,
+            seed=args.seed,
+            patience=args.patience,
+            project=str(project.resolve()),
+            name=run_name,
+            exist_ok=True,
+            pretrained=True,
+            cos_lr=True,
+            amp=True,
+            verbose=True,
+        )
+
+        best_path = Path(train_results.save_dir) / "weights" / "best.pt"
+        roi_save_dir = Path(train_results.save_dir) / "roi_eval"
+        roi_summary = evaluate_roi(best_path, roi_dir, roi_save_dir, args.device)
+
+        record = {
+            "model": model_name,
+            "run_name": run_name,
+            "save_dir": str(Path(train_results.save_dir).resolve()),
+            "best_pt": str(best_path.resolve()),
+            "metrics": {
+                "map50": float(train_results.results_dict.get("metrics/mAP50(B)", 0.0)),
+                "map50_95": float(train_results.results_dict.get("metrics/mAP50-95(B)", 0.0)),
+                "precision": float(train_results.results_dict.get("metrics/precision(B)", 0.0)),
+                "recall": float(train_results.results_dict.get("metrics/recall(B)", 0.0)),
+            },
+            "roi_eval": roi_summary,
+        }
+        compare_summary.append(record)
+
+        summary_path = Path(train_results.save_dir) / "roi_eval_summary.json"
+        summary_path.write_text(json.dumps(record, indent=2, ensure_ascii=False), encoding="utf-8")
+
+        compare_json = project / "compare_summary.json"
+        compare_json.write_text(json.dumps(compare_summary, indent=2, ensure_ascii=False), encoding="utf-8")
+
+        print(json.dumps(record, indent=2, ensure_ascii=False))
+
+
+if __name__ == "__main__":
+    main()
--- a/10_run_shoe_compare.ps1
+++ b/10_run_shoe_compare.ps1
@ -0,0 +1,19 @@
+$ErrorActionPreference = "Stop"
+
+$repo = "C:\Users\tianj\Documents\apps\DetectionModelTraining"
+$env:PYTHONPATH = "$repo\.pydeps"
+$env:YOLO_CONFIG_DIR = "$repo\.ultralytics"
+
+Set-Location $repo
+
+py -3.11 .\07_build_public_shoe_dataset.py --clean
+
+py -3.11 .\08_train_compare_models.py `
+  --data "$repo\datasets\shoe-public-mix\data.yaml" `
+  --roi-dir "$repo\datasets\roi-shoes" `
+  --project "$repo\runs\shoe_compare" `
+  --models "yolov8s.pt" "yolo11s.pt" "yolo26s.pt" `
+  --epochs 40 `
+  --imgsz 640 `
+  --batch 16 `
+  --device 0
--- a/11_run_shoe_compare_960.ps1
+++ b/11_run_shoe_compare_960.ps1
@ -0,0 +1,17 @@
+$ErrorActionPreference = "Stop"
+
+$repo = "C:\Users\tianj\Documents\apps\DetectionModelTraining"
+$env:PYTHONPATH = "$repo\.pydeps"
+$env:YOLO_CONFIG_DIR = "$repo\.ultralytics"
+
+Set-Location $repo
+
+py -3.11 .\08_train_compare_models.py `
+  --data "$repo\datasets\shoe-public-mix\data.yaml" `
+  --roi-dir "$repo\datasets\roi-shoes" `
+  --project "$repo\runs\shoe_compare_960" `
+  --models "yolov8s.pt" "yolo11s.pt" "yolo26s.pt" `
+  --epochs 40 `
+  --imgsz 960 `
+  --batch 8 `
+  --device 0
--- a/README.md
+++ b/README.md
@ -0,0 +1,156 @@
+# 鞋子检测模型训练指南
+
+## 方案：640x640 单模型（部署时用2窗口）
+
+**训练阶段**：
+- 输入：640x640 完整图片
+- 模型：YOLOv8s
+- 输出：640x640 模型文件
+
+**部署阶段**（pipeline配置）：
+- 原图 1920x1080
+- 分成 2 个 960x1080 窗口
+- 每个窗口 resize 到 640x640 送入模型
+- 合并检测结果
+
+---
+
+## 目录结构
+
+```
+train/
+├── README.md                 # 本文件
+├── 01_download_dataset.py    # 下载鞋子数据集（推荐 Open Images）
+├── 02_train.bat              # Windows 一键训练脚本
+├── 03_export_onnx.bat        # 导出 ONNX 脚本
+├── 04_convert_rknn.py        # 转换为 RKNN 脚本
+├── 05_prepare_ppe_shoe_subset.py # 提取 PPE 鞋子单类子集
+├── 06_finetune_ppe.bat       # 用 PPE 鞋子子集做二阶段微调
+├── data.yaml.template        # 数据集配置文件
+└── samples/                  # 示例图片
+    ├── calibration/          
+    ├── test_images/          
+    └── README.md
+```
+
+---
+
+## 快速开始
+
+### 1. 下载数据集
+
+```bash
+cd train
+python 01_download_dataset.py --source openimages --max-samples 5000
+```
+
+### 2. 准备配置
+
+```bash
+脚本会自动生成 datasets/openimages-shoes-yolo/data.yaml
+```
+
+### 3. 训练（640x640）
+
+```bash
+02_train.bat
+```
+
+或手动：
+```bash
+yolo detect train \
+    data=datasets/openimages-shoes-yolo/data.yaml \
+    model=yolov8s.pt \
+    epochs=150 \
+    imgsz=640 \
+    batch=16 \
+    device=0
+```
+
+**训练参数**：
+- 模型：YOLOv8s（速度和精度平衡）
+- 输入：640x640
+- 预计时间：30-60分钟
+
+### 4. 导出 ONNX
+
+```bash
+03_export_onnx.bat
+```
+
+### 5. 转换为 RKNN
+
+在 Ubuntu PC 上：
+```bash
+python 04_convert_rknn.py runs/detect/train/weights/best.onnx -o shoe_detector_640.rknn -t rk3588
+```
+
+### 6. 部署（2窗口配置）
+
+复制到 RK3588：
+```bash
+scp shoe_detector_640.rknn orangepi@<rk3588_ip>:/home/orangepi/apps/OrangePi3588Media/models/
+```
+
+Pipeline 配置（部署阶段用2窗口）：
+```json
+{
+  "id": "pre_shoe",
+  "type": "preprocess",
+  "windows": [
+    {"x": 0, "y": 0, "w": 960, "h": 1080},
+    {"x": 960, "y": 0, "w": 960, "h": 1080}
+  ],
+  "dst_w": 640,
+  "dst_h": 640
+}
+```
+
+### 7. 方案 A：PPE 二阶段微调
+
+当 Open Images 基础模型训练完成后，可继续用 PPE 鞋子子集做场景微调：
+
+```bash
+python 05_prepare_ppe_shoe_subset.py
+06_finetune_ppe.bat
+```
+
+PPE 鞋子子集来源：
+- `boots`
+- `no_boots`
+
+这两个类会统一映射成单类：
+- `shoe`
+
+---
+
+## 类别说明（Open Images）
+
+Open Images 官方鞋类层级中，`Footwear` 的子类包括：
+- `Boot`
+- `Sandal`
+- `High heels`
+- `Roller skates`
+
+本项目推荐下载：
+- `Footwear`
+- `Boot`
+
+可选补充：
+- `Sandal`
+
+不建议默认加入：
+- `High heels`
+- `Roller skates`
+
+训练时统一映射为单一类别：
+- `0: shoe`
+
+这样模型目标更聚焦，先尽量把鞋子稳定检出，再在后处理里判断是否为黑色鞋。
+
+---
+
+## 相关链接
+
+- [Open Images 数据集](https://storage.googleapis.com/openimages/web/index.html)
+- [Ultralytics YOLOv8](https://docs.ultralytics.com/)
--- a/data.yaml.template
+++ b/data.yaml.template
@ -0,0 +1,20 @@
+# 单类鞋子检测数据集配置
+
+path: .
+train: images/train
+val: images/val
+test: images/test
+
+# 单类 shoe
+nc: 1
+names: ['shoe']
+
+# 推荐来源:
+#   python 01_download_dataset.py --source openimages --max-samples 5000
+#
+# 推荐 Open Images 子类:
+#   - Footwear
+#   - Boot
+#
+# 可选补充:
+#   - Sandal
--- a/samples/README.md
+++ b/samples/README.md
@ -0,0 +1,46 @@
+# 示例图片目录
+
+用于存放测试图片和量化校准样本。
+
+## 目录结构
+
+```
+samples/
+├── test_images/      # 用于测试模型的示例图片
+├── calibration/      # INT8 量化校准用的图片（约 20-100 张）
+└── README.md         # 本文件
+```
+
+## 使用说明
+
+### 测试图片 (test_images/)
+
+存放一些典型场景的鞋子图片，用于验证模型效果。
+
+### 校准图片 (calibration/)
+
+INT8 量化时需要，用于确定量化参数。
+
+**要求：**
+- 应与实际部署场景相似
+- 包含各种光照、角度、背景的样本
+- 建议 20-100 张
+- 图片格式: JPG, PNG, BMP
+
+**创建校准数据集文件：**
+
+```bash
+# Linux/macOS
+ls samples/calibration/*.jpg > dataset.txt
+ls samples/calibration/*.png >> dataset.txt
+
+# Windows CMD
+dir /b samples\calibration\*.jpg > dataset.txt
+dir /b samples\calibration\*.png >> dataset.txt
+```
+
+## 注意事项
+
+- 校准图片越多，转换时间越长，但精度可能更高
+- 建议使用训练集的部分图片作为校准集
+- 不要和测试集重复，避免过拟合