🎓 Educational Refactoring | 教学导向重构 A modern PyTorch implementation of the CVPR 2015 paper with comprehensive tutorials and documentation. 使用现代 PyTorch 框架实现 CVPR 2015 论文，包含完整教程和文档。
This project is a complete refactoring of the classic Person Re-Identification (ReID) algorithm from the CVPR 2015 paper:
本项目是经典人员重识别算法的完全重构，基于 CVPR 2015 论文：
"An Improved Deep Learning Architecture for Person Re-Identification" by Ejaz Ahmed, Michael Jones, and Tim K. Marks
Paper: CVPR 2015 PDF
- ✅ Modern Tech Stack: Migrate from TensorFlow 1.x to PyTorch 2.5+ / 使用 PyTorch 2.5+ 替代 TensorFlow 1.x
- ✅ Educational Focus: Comprehensive Jupyter notebooks and bilingual docs / 完整教学笔记本和中英文档
- ✅ Best Practices: PyTorch Lightning, Type Hints, Modular Design / Lightning 框架、类型提示、模块化设计
- ✅ Reproducibility: uv package manager, YAML configs / uv 包管理、YAML 配置
# Clone repository | 克隆仓库
git clone https://github.com/Ning-Ding/Implementation-CVPR2015-CNN-for-ReID.git
cd Implementation-CVPR2015-CNN-for-ReID
# Install dependencies with uv | 使用 uv 安装依赖
uv sync --all-extras
# Activate virtual environment | 激活虚拟环境
source .venv/bin/activate # Linux/Mac
# or
.venv\\Scripts\\activate # Windows# Train on CUHK03 dataset | 在 CUHK03 数据集上训练
python scripts/train.py --config config/cuhk03.yaml
# Train on Market-1501 dataset | 在 Market-1501 数据集上训练
python scripts/train.py --config config/market1501.yamljupyter lab
# Open notebooks in order:
# 1. notebooks/01_introduction.ipynb
# 2. notebooks/02_data_exploration.ipynb
# 3. notebooks/03_model_architecture.ipynb
# 4. notebooks/04_training_demo.ipynb
# 5. notebooks/05_evaluation.ipynb
# 6. notebooks/06_inference_demo.ipynbImplementation-CVPR2015-CNN-for-ReID/
├── src/ # 核心代码 | Core Code
│ ├── data/ # 数据集和数据增强 | Datasets & Augmentation
│ │ ├── base_dataset.py
│ │ ├── cuhk03_dataset.py
│ │ ├── market1501_dataset.py
│ │ └── transforms.py
│ ├── models/ # 模型定义 | Model Definitions
│ │ ├── layers.py # 自定义层 (Cross-Input) | Custom Layers
│ │ ├── siamese_cnn.py # 原始论文架构 | Original Architecture
│ │ └── lightning_module.py
│ ├── training/ # 训练逻辑 | Training Logic
│ ├── evaluation/ # 评估指标 (CMC, mAP) | Metrics
│ └── utils/ # 工具函数 | Utilities
│
├── config/ # YAML 配置文件 | Config Files
│ ├── base.yaml
│ ├── cuhk03.yaml
│ └── market1501.yaml
│
├── notebooks/ # 📓 Jupyter 教学笔记本 | Tutorials
│ ├── 01_introduction.ipynb
│ ├── 02_data_exploration.ipynb
│ ├── 03_model_architecture.ipynb
│ ├── 04_training_demo.ipynb
│ ├── 05_evaluation.ipynb
│ └── 06_inference_demo.ipynb
│
├── docs/ # 📚 技术文档 | Documentation
│ ├── paper_theory.md # 论文原理详解
│ ├── architecture.md # 网络架构详解
│ └── algorithm_explanation.md
│
├── scripts/ # 可执行脚本 | Executable Scripts
│ ├── train.py
│ └── evaluate.py
│
├── legacy/ # 旧代码参考 | Legacy Code
│ ├── CUHK03/main.py
│ └── market1501/
│
├── pyproject.toml # uv 项目配置 | Project Config
└── README.md # 本文件 | This File
The Siamese CNN consists of: 孪生 CNN 网络包含：
-
Tied Convolution Layers (权重共享卷积层)
- Conv1: 20 filters @ 5×5 → MaxPool 2×2
- Conv2: 25 filters @ 5×5 → MaxPool 2×2
-
Cross-Input Neighborhood Differences (交叉输入邻域差异) 🌟
- Paper's core innovation | 论文核心创新
- Captures local spatial relationships between image pairs
- 捕获图像对之间的局部空间关系
-
Patch Summary Features (图块摘要特征)
- Conv 5×5 with stride=5
-
Across-Patch Features (跨图块特征)
- Conv 3×3 → MaxPool 2×2
-
Fully Connected Layers (全连接层)
- FC1: 4250 → 500 (with ReLU)
- FC2: 500 → 2 (same/different person)
- Identities: 1,467 (1,360 labeled + 107 detected)
- Images: ~14,000
- Cameras: 5 pairs
- Download: CUHK03 Official
- Identities: 1,501
- Images: 32,668 (12,936 train + 19,732 gallery/query)
- Cameras: 6
- Download: Market-1501 Dataset
| Component | Technology |
|---|---|
| Deep Learning | PyTorch 2.5+, PyTorch Lightning 2.4+ |
| Data Processing | NumPy 2.0+, Albumentations 2.0+ |
| Package Manager | uv (modern pip replacement) |
| Configuration | YAML, OmegaConf |
| Logging | TensorBoard, Weights & Biases (optional) |
| Visualization | Matplotlib, Seaborn, Plotly |
| Code Quality | Black, Ruff, MyPy, Pytest |
model:
input_size: [160, 60]
loss:
type: contrastive
margin: 2.0
optimizer:
name: sgd
lr: 0.01
momentum: 0.9
weight_decay: 0.00025
scheduler:
name: polynomial
power: 0.75
training:
max_epochs: 2000
batch_size: 150- ✅ Cross Entropy: Binary classification (same/different)
- ✅ Contrastive Loss: Siamese networks (recommended)
- ✅ Triplet Loss: Margin-based learning
| Metric | Description | 说明 |
|---|---|---|
| CMC | Cumulative Matching Characteristic | 累积匹配特征曲线 |
| mAP | mean Average Precision | 平均精度均值 |
| Rank-1 | Top-1 accuracy | Top-1 准确率 |
| Rank-5/10/20 | Top-K accuracy | Top-K 准确率 |
Example Results on CUHK03:
Rank-1: ~62%
Rank-5: ~84%
Rank-10: ~91%
mAP: ~60%
- ❌ TensorFlow 1.x → ✅ PyTorch 2.5+
- ❌ Python 2 syntax → ✅ Python 3.11+ with Type Hints
- ❌ Hardcoded paths → ✅ YAML configuration
- ❌ No tests → ✅ Pytest test suite
- ❌ Scattered code → ✅ Modular design
- ✅ PyTorch Lightning (简化训练流程)
- ✅ Albumentations (高性能数据增强)
- ✅ uv package manager (快速依赖管理)
- ✅ Rich logging (美观的终端输出)
- ✅ Type hints (类型安全)
- ✅ Comprehensive documentation (完整文档)
- Introduction - ReID 问题背景与应用场景
- Data Exploration - 数据集统计分析和可视化
- Model Architecture - 网络结构逐层讲解（含公式推导）
- Training Demo - 训练流程演示与监控
- Evaluation - CMC/mAP 指标详解
- Inference - 模型部署与推理优化
Contributions are welcome! Please:
- Fork the repository
- Create a feature branch
- Follow code style (Black + Ruff)
- Add tests for new features
- Submit a pull request
This project is licensed under the MIT License - see the LICENSE file for details.
Original implementation by Ning Ding (2017) Refactored with modern frameworks (2024)
- Original paper authors: Ejaz Ahmed, Michael Jones, Tim K. Marks
- Original TensorFlow implementation: Ning Ding
- CUHK03 dataset creators
- Market-1501 dataset creators
For questions or discussions: 如有问题或讨论：
- Open an issue on GitHub
- Original author: [email protected]
If you find this project helpful, please consider giving it a star ⭐!
如果这个项目对你有帮助，请给个星标支持！
Last Updated: 2024 Status: ✅ Active Development