tiny-detr 🔍

A lightweight implementation of the DETR (DEtection TRansformer) object detection model, using ResNet-18 as the backbone and a simplified transformer architecture.

🌟 Highlights

• Lightweight: Only ~13.9M parameters (compared to original DETR's 41M+)
• Fast Training: 2-3x faster training compared to original DETR
• Simple Architecture: Perfect for learning and experimentation
• Multi-Dataset Support: Works with COCO and PASCAL VOC

🏗️ Architecture

Component	Specification
Backbone	ResNet-18
Transformer Encoder	2 layers, 4 heads
Transformer Decoder	2 layers, 4 heads
Hidden Dimension	128
Object Queries	100

📈 Performance

Metric	Value
mAP	17.40
mAP@50	39.14
mAP@75	14.02

🚀 Quick Start

Installation

# Clone the repository
git clone https://github.com/ilyasoulk/tiny-detr.git
cd tiny-detr/src/detr

# Create virtual environment (optional but recommended)
python -m venv venv
source venv/bin/activate  # Linux/Mac
# or
.\venv\Scripts\activate  # Windows

# Install dependencies
pip install -r requirements.txt

Training

python src/detr/train.py

Inference

python src/detr/inference.py

📁 Project Structure

src/detr/
├── train.py          # Training script
├── inference.py      # Inference and visualization
├── model.py          # Model architecture
├── loss.py          # Bipartite matching loss
├── metrics.py        # mAP calculation
├── engine.py         # Training and evaluation loops
├── dataset.py        # COCO dataset handling
└── dataset_voc.py    # VOC dataset handling

⚙️ Configuration

config = {
    'backbone': 'resnet18',
    'hidden_dim': 128,
    'num_heads': 4,
    'num_encoder': 2,
    'num_decoder': 2,
    'num_queries': 100,
    'learning_rate': 1e-4,
    'weight_decay': 1e-4
}

📊 Dataset Support

Supported datasets:

• COCO: Full support with automatic download
• PASCAL VOC: Full support with automatic download

🙏 Acknowledgments

This implementation is inspired by the original DETR paper:

• End-to-End Object Detection with Transformers

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.