This repository is an official implementation of the Anchor DETR. We encode the anchor points as the object queries in DETR. Multiple patterns are attached to each anchor point to solve the difficulty: "one region, multiple objects". We also propose an attention variant RCDA to reduce the memory cost for high-resolution features.

Note:

1. The results are based on ResNet-50 backbone.
2. Inference speeds are measured on NVIDIA Tesla V100 GPU.
3. DETR gets 12 FPS with torchscript optimization and 10 FPS without it. The others do not enable the torchscript.

Note: the models and logs are also available at Baidu Netdisk with code f56r.

First, clone the repository locally:

git clone https://github.com/megvii-research/AnchorDETR.git

Then, install dependencies:

pip install -r requirements.txt

To train AnchorDETR on a single node with 8 GPUs:

python -m torch.distributed.launch --nproc_per_node=8 --use_env main.py  --coco_path /path/to/coco 

To evaluate AnchorDETR on a single node with 8 GPUs:

python -m torch.distributed.launch --nproc_per_node=8 --use_env main.py --eval --coco_path /path/to/coco --resume /path/to/checkpoint.pth 

To evaluate AnchorDETR with a single GPU:

python main.py --eval --coco_path /path/to/coco --resume /path/to/checkpoint.pth

If you find this project useful for your research, please consider citing the paper.

@misc{wang2021anchor,
      title={Anchor DETR: Query Design for Transformer-Based Detector},
      author={Yingming Wang and Xiangyu Zhang and Tong Yang and Jian Sun},
      year={2021},
      eprint={2109.07107},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

If you have any questions, feel free to open an issue or contact us at [email protected].