Align-DETR: Improving DETR with Simple IoU-aware BCE loss

Zhi Cai, Songtao Liu, Guodong Wang, Zheng Ge, Xiangyu zhang, Di Huang*

Arxiv version link

Introduction

AlignDETR is a variant of DETR(DEtection with Transformer), with a simple IoU-Aware BCE loss and better performance! It aims to solve the issue of misalignment problem spotted in DETR's output.

News 📰

[2024.07.21] We release a stronger version of Align-DETR by making a smooth modification on IA-BCE loss (Now we term as Align Loss in the newest BMVC version)

[2024.07.21] Align-DETR is accepted by BMVC!

Installation 🛠️

Install details can be found in installation instructions

Usage 📖

Train Example

python tools/train_net.py --config-file  aligndetr/aligndetr_k=2_r50_4scale_12ep.py --num-gpus 8

Evaluation Example

python tools/train_net.py --config-file  aligndetr/aligndetr_k=2_r50_4scale_12ep.py --num-gpus 8 --eval train.init_checkpoint=/path/to/checkpoint

Model Zoo 🦁

* represents using a modified IA-BCE loss that absorbs focal loss term.

Model	AP	AP50	AP75	APs	APm	APl	weights
AlignDETR-R50-12ep	50.3	67.9	54.8	34.1	53.5	65.1	Google Drive
AlignDETR-R50-24ep	51.4	69.1	55.8	35.5	54.6	65.7	Google Drive
AlignDETR-R50-12ep*	50.5	67.7	55.3	34.7	53.6	64.6	Google Drive
AlignDETR-R50-24ep*	51.7	69.0	56.3	35.5	55.0	66.1	Google Drive

Acknowlegements 🙏

1. Our code is based on detrex and detectron2.
2. Align-DETR is also available in the open-source benchmark detrex and mmdetection now!

Citation

If you are interested in our work and use our method in your research, please cite

@misc{cai2023aligndetr,
      title={Align-DETR: Improving DETR with Simple IoU-aware BCE loss}, 
      author={Zhi Cai and Songtao Liu and Guodong Wang and Zheng Ge and Xiangyu Zhang and Di Huang},
      year={2023},
      eprint={2304.07527},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

License

This project is released under the Apache 2.0 license.