🔍 Person Re-Identification System

A Siamese Neural Network-based person re-identification system comparing Triplet Loss vs Contrastive Loss.

📁 Project Structure

Siemese_ATML/
├── data/                           # Dataset
│   └── SNN-TL-Data/
│       ├── train/                  # Training images
│       └── train.csv               # Triplet annotations
├── models/                         # Trained models
│   ├── triplet/                    # Triplet Loss model
│   │   ├── best_model.pt
│   │   └── database.csv
│   └── contrastive/                # Contrastive Loss model
│       ├── best_model_contrastive.pt
│       └── database_contrastive.csv
├── outputs/                        # Generated outputs
│   ├── training_history.png
│   ├── training_history_contrastive.png
│   ├── triplet_sample.png
│   └── results.png
├── src/                            # Training scripts
│   ├── train_triplet.py
│   └── train_contrastive.py
├── archive/                        # Legacy files
├── app.py                          # Streamlit web app
├── requirements.txt                # Dependencies
└── README.md

🚀 Quick Start

1. Install Dependencies

pip install -r requirements.txt

2. Run Streamlit App

streamlit run app.py

Open http://localhost:8501 in your browser.

🧠 Models

Loss Function	Description	Model File
Triplet Loss	Uses anchor-positive-negative triplets	models/triplet/best_model.pt
Contrastive Loss	Uses positive-negative pairs	models/contrastive/best_model_contrastive.pt

Both models use EfficientNet-B0 backbone with 512-dimensional embeddings.

📊 Training

Train Triplet Loss Model

python src/train_triplet.py

Train Contrastive Loss Model

python src/train_contrastive.py

🖥️ Features

• Model Selection: Switch between Triplet and Contrastive loss models
• Compare Mode: View results from both models side-by-side
• GPU Accelerated: Automatic CUDA support with mixed precision (AMP)
• Interactive UI: Upload images and find matching persons

⚙️ Configuration

Training parameters can be modified in src/train_*.py:

• BATCH_SIZE: 64 (default)
• EPOCHS: 10
• LR: 0.001
• EARLY_STOP_PATIENCE: 3

📜 License

MIT License