🚧 REG: Refined Generalized Focal Loss

🌟 Highlights

• ✅ REG Loss: Combines focal loss, spatial refinement, and uncertainty.
• 🎯 Multi-task Model: Detection + segmentation in one architecture.
• 📊 Performance: mAP₅₀ = 80.34, F1 = 77.87
• 🧱 Modular PyTorch: Easy to adapt and extend.
• 📦 Dockerized: CVPR-ready reproducibility.

🗂 Project Structure

REG/
├── src/
│   ├── model.py         # REG Loss + Multi-task Detection-Segmentation Model
│   ├── train.py         # Training Pipeline
│   ├── inference.py     # Inference Script
│   ├── metrics.py       # Evaluation Metrics
│   └── utils.py         # Utility Functions
├── Dockerfile
├── requirements.txt
└── README.md

🚀 Getting Started

🔧 1. Install

git clone https://github.com/kaopanboonyuen/REG.git
cd REG
pip install -r requirements.txt

🐳 Docker

docker build -t reg-road-assets .

📊 Training

python src/train.py \
  --data-path ./data/ \
  --batch 8 \
  --epochs 50 \
  --lr 1e-4

🖼 Inference

python src/inference.py \
  --checkpoint ./checkpoints/best_model.pth \
  --image-path ./images/test.jpg \
  --output-path ./output/result.png

📏 Evaluation

python src/metrics.py \
  --pred-dir ./predictions/ \
  --gt-dir ./ground_truth/

📈 Results Summary

📚 Learn More

• 📖 Blog: REG Explained
• 📄 Paper on arXiv
• 📄 Paper on IEEE

📄 Citation

@inproceedings{panboonyuen2025reg,
  title={REG: Refined Generalized Focal Loss for Road Asset Detection on Thai Highways Using Vision Models},
  author={Panboonyuen, Teerapong},
  booktitle={2025 17th International Conference on Knowledge and Smart Technology (KST)},
  pages={324--329},
  year={2025},
  organization={IEEE}
}

👨‍💻 Contact

Teerapong Panboonyuen (Kao)
Postdoctoral Researcher, Chulalongkorn University
Senior Research Scientist, MARSAIL
📧 teerapong.panboonyuen@gmail.com

🤝 Contributing

Pull requests are welcome! Open an issue to propose changes or improvements.

🧠 License

MIT License — free to use, share, and modify.