UniHead: an effective detection head

📢 Introduction

This is the official implementation of our paper titled "UniHead: Unifying Multi-Perception for Detection Heads".

For more details, please refer to our paper. This repo is based on PyTorch>=1.7.1 and mmdet==2.25.1.

🏆 Result

Our UniHead can bring significant AP improvements to a large number of detectors, even with lower model complexity.

📂 Weight

We release the weight of models used in our paper, including RetinaNet, FCOS, RepPoints, FreeAnchor, ATSS and GFL. You may need to log out of your Google account to download them.

Download the weight(s) from corresponding links below.

• RetinaNet-UniHead (39.2 AP): Google Drive; BaiduNetDisk

• FCOS-UniHead (40.4 AP): Google Drive; BaiduNetDisk

• RepPoints-UniHead (39.9 AP): Google Drive; BaiduNetDisk

• FreeAnchor-UniHead (41.6 AP): Google Drive; BaiduNetDisk

• ATSS-UniHead (41.2 AP): Google Drive; BaiduNetDisk

• GFL-UniHead (42.3 AP): Google Drive; BaiduNetDisk

The weights of UniHead with various of backbones (R101 and SwinT) are also released. If you need more models, such as UniHead-SwinB, please feel free to contact me via email.

• UniHead-R101-2x-MS (47.7 AP): Google Drive; BaiduNetDisk

• UniHead-SwinT-3x-MS (51.3 AP): Google Drive; BaiduNetDisk

🚀 Training

More detail please see mmdetection.

bash dist_train_UniHead.sh

🎡 Visualization

Examples of detection results obtained by our UniHead and baseline (RetinaNet).

📚 Citation

Please cite our work if you find our work and codes helpful for your research.

@article{zhou2023unihead,
  title={UniHead: Unifying Multi-Perception for Detection Heads},
  author={Zhou, Hantao and Yang, Rui and Zhang, Yachao and Duan, Haoran and Huang, Yawen and Hu, Runze and Li, Xiu and Zheng, Yefeng},
  journal={arXiv preprint arXiv:2309.13242},
  year={2023}
}

Acknowledgement

This project is built upon numerous previous projects. We'd like to thank the contributors of mmdetection.