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.
Install details can be found in installation instructions
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
* 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 |
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}
}
This project is released under the Apache 2.0 license.