Rank-DETR for High Quality Object Detection (NeurIPS 2023)

Yifan Pu, Weicong Liang, Yiduo Hao, Yuhui Yuan, Yukang Yang, Chao Zhang, Han Hu, and Gao Huang

[arXiv] [BibTeX]

Table of Contents

• Installation
• Pretrained Models
• Run
  • Training
  • Evaluation
• Citation

Installation

Please refer to the installation document of detrex.

Pretrained Models

Here we provide the Rank-DETR model pretrained weights based on detrex:

Name	Backbone	Query Num	Epochs	AP	download
Rank-DETR	R50	300	12	50.2	model
Rank-DETR	R50	300	36	51.2	model
Rank-DETR	Swin Tiny	300	12	52.7	model
Rank-DETR	Swin Tiny	300	36	54.7	model
Rank-DETR	Swin Large	300	12	57.3	model
Rank-DETR	Swin Large	300	36	58.2	model

Run

Training

All configs can be trained with:

cd detrex
python projects/rank_detr/train_net.py --config-file projects/rank_detr/configs/path/to/config.py --num-gpus 8

• By default, we use 8 GPUs with total batch size as 16 for training.
• To train/eval a model with the swin transformer backbone, you need to download the backbone from the offical repo frist and specify argument train.init_checkpoint like our configs.

Evaluation

Model evaluation can be done as follows:

cd detrex
python projects/rank_detr/train_net.py --config-file projects/rank_detr/configs/path/to/config.py  --eval-only train.init_checkpoint=/path/to/model_checkpoint

Citing Rank-DETR

If you find Rank-DETR useful in your research, please consider citing:

@inproceedings{pu2023rank,
  title={Rank-DETR for High Quality Object Detection},
  author={Pu, Yifan and Liang, Weicong and Hao, Yiduo and Yuan, Yuhui and Yang, Yukang and Zhang, Chao and Hu, Han and Huang, Gao},
  booktitle={NeurIPS},
  year={2023}
}