Anchor DETR: Query Design for Transformer-Based Detector

Introduction

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.

DETR

Main Results

	feature	epochs	AP	GFLOPs	Infer Speed (FPS)
DETR	DC5	500	43.3	187	10 (12)
SMCA	multi-level	50	43.7	152	10
Deformable DETR	multi-level	50	43.8	173	15
Conditional DETR	DC5	50	43.8	195	10
Anchor DETR	DC5	50	44.3	172	16 (19)

Note:

The results are based on ResNet-50 backbone.
Inference speeds are measured on NVIDIA Tesla V100 GPU.
The values in parentheses of the Infer Speed indicate the speed with torchscript optimization.

Model

name	backbone	AP	URL
AnchorDETR-C5	R50	42.1	model / log
AnchorDETR-DC5	R50	44.3	model / log
AnchorDETR-C5	R101	43.5	model / log
AnchorDETR-DC5	R101	45.1	model / log

Note: the models and logs are also available at Baidu Netdisk with code hh13.

Usage

Installation

First, clone the repository locally:

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

Then, install dependencies:

pip install -r requirements.txt

Training

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

Evaluation

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

Citation

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

@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}
}

Contact

If you have any questions, feel free to open an issue or contact us at wangyingming@megvii.com.

megvii-research / AnchorDETR

readme