projects/anchor_detr/README.md

Anchor DETR: Query Design for Transformer-Based Object Detection

Yingming Wang, Xiangyu Zhang, Tong Yang, Jian Sun

[arXiv] [BibTeX]

Pretrained Weights

Here's our pretrained Anchor-DETR weights based on detrex.

Name	Backbone	Pretrain	Epochs	box AP	download
Anchor-DETR-R50	R-50	IN1k	50	41.9	model

Converted Weights

Name	Backbone	Pretrain	Epochs	box AP	download
Anchor-DETR-R50	R-50	IN1k	50	42.2	model
Anchor-DETR-R50-DC5	R-50	IN1k	50	44.2	model
Anchor-DETR-R101	R-101	IN1k	50	43.5	model
Anchor-DETR-R101-DC5	R-101	IN1k	50	45.1	model

Note: Here we borrowed the pretrained weight from Anchor-DETR official repo. And our detrex training results will be released in the future version.

Training

Training Anchor-DETR-R50 model:

cd detrex
python tools/train_net.py --config-file projects/anchor_detr/configs/anchor_detr_r50_50ep.py --num-gpus 8

By default, we use 8 GPUs with total batch size as 64 for training.

Evaluation

Model evaluation can be done as follows:

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

Citing Anchor-DETR

@inproceedings{wang2022anchor,
  title={Anchor detr: Query design for transformer-based detector},
  author={Wang, Yingming and Zhang, Xiangyu and Yang, Tong and Sun, Jian},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={36},
  number={3},
  pages={2567--2575},
  year={2022}
}