Gait Biometric Identification System

University Undergraduate Thesis:

A state-of-the-art gait biometric identification system with a Gradient Reversal Layer (GRL) for view-invariant learning.

Features

• ✨ Architecture: Set-based deep learning for gait recognition
• 🔄 Gradient Reversal Layer: Domain adaptation for view-invariant features
• 🚀 Multi-Device Support: CUDA, MPS (Apple Silicon), and CPU
• 📊 Comprehensive Evaluation: Rank-k accuracy, mAP, CMC curves
• ⚙️ Flexible Configuration: YAML-based configuration system
• 📈 TensorBoard Integration: Real-time training monitoring

Quick Start

Installation

# Install dependencies
pip install -r requirements.txt

Dataset Preparation

Download CASIA-B dataset and organize it as:

CASIA-B/casiab-128-end2end/
├── 001/
│   ├── nm-01/
│   │   ├── 000/
│   │   │   └── 000-sils.pkl
│   │   └── ...
│   └── ...
└── ...

Update the dataset path in configs/config.yaml:

dataset:
  data_root: "/path/to/CASIA-B/casiab-128-end2end"

Training

# Train with GRL enabled
python train.py --config configs/config.yaml

# Or use the shell script
cd scripts
./train.sh

Evaluation

# Evaluate trained model
python scripts/evaluate.py \
    --config configs/config.yaml \
    --checkpoint output/best_model.pth \
    --output_dir evaluation_results \
    --visualize

Monitoring

# Launch TensorBoard
tensorboard --logdir output/tensorboard

Project Structure

gait_biometric_identification/
├── configs/              # Configuration files
├── data/                 # Dataset loaders and transforms
├── models/               # Model architectures and losses
├── utils/                # Utilities (metrics, visualization, device)
├── scripts/              # Training and evaluation scripts
├── train.py              # Main training script
├── requirements.txt      # Dependencies
├── wiki.md              # Comprehensive documentation
└── README.md            # This file

Configuration

Enable/Disable GRL

In configs/config.yaml:

model:
  grl:
    enabled: true  # Set to false to disable GRL
    lambda_grl: 1.0
    schedule: "constant"  # or "progressive"

Device Selection

device:
  type: "cuda"  # Options: "cuda", "mps", "cpu"
  gpu_ids: [0]

Training Parameters

training:
  batch_size: 8
  person_num: 8  # P persons per batch
  sample_num: 16  # K samples per person
  num_epochs: 200
  
  optimizer:
    lr: 0.0001

Results

Best performance on CASIA-B:

Setting	Rank-1	Rank-5	mAP
With GRL	98.95%	99.83%	82.16%
Without GRL	98.78%	99.48%	79.99%

Note: Results vary based on training settings and random seed

Key Components

1. Backbone

• Set-based feature extraction
• Horizontal Pyramid Pooling
• Temporal aggregation (max + mean)

2. Gradient Reversal Layer (GRL)

• Domain adaptation for view angles
• Adversarial training
• Configurable lambda scheduling

3. Loss Functions

• Identity classification loss
• Triplet loss (batch hard mining)
• Center loss (intra-class compactness)
• View classification loss (for GRL)

4. Evaluation Metrics

• Rank-k accuracy (k=1, 5, 10)
• Mean Average Precision (mAP)
• Cumulative Match Characteristic (CMC)

Documentation

See wiki.md for comprehensive documentation including:

• Detailed architecture explanation
• Mathematical formulations
• Training pipeline details
• Troubleshooting guide
• Advanced topics

Requirements

• Python >= 3.8
• PyTorch >= 2.0.0
• CUDA >= 11.0 (for GPU training)
• See requirements.txt for complete list

License

This project is for academic research purposes.

Acknowledgments

• CASIA-B dataset providers
• PyTorch team

Contact

For questions or issues, please refer to the wiki.md troubleshooting section.