H-PETR-Pose

This is the official implementation of the paper "DETRs with Hybrid Matching".

Authors: Ding Jia, Yuhui Yuan, Haodi He, Xiaopei Wu, Haojun Yu, Weihong Lin, Lei Sun, Chao Zhang, Han Hu

Citing H-PETR-Pose

If you find H-PETR-Pose useful in your research, please consider citing:

@article{jia2022detrs,
  title={DETRs with Hybrid Matching},
  author={Jia, Ding and Yuan, Yuhui and He, Haodi and Wu, Xiaopei and Yu, Haojun and Lin, Weihong and Sun, Lei and Zhang, Chao and Hu, Han},
  journal={arXiv preprint arXiv:2207.13080},
  year={2022}
}

@inproceedings{shi2022end,
  title={End-to-End Multi-Person Pose Estimation With Transformers},
  author={Shi, Dahu and Wei, Xing and Li, Liangqi and Ren, Ye and Tan, Wenming},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={11069--11078},
  year={2022}
}

Model ZOO

We provide a set of baseline results and trained models available for download:

Name	Backbone	epochs	AP (Reproduced / Reported)	download
Deformable-DETR	R50	100	69.3 / 68.8	model
Deformable-DETR	R101	100	69.9 / 70.0	model
Deformable-DETR	Swin Large	100	73.3 / 73.1	model
H-Deformable-DETR	R50	100	70.9	model
H-Deformable-DETR	R101	100	71.0	model
H-Deformable-DETR	Swin Large	100	74.9	model

• We use 8 V-100 GPUs and batch_size = 8 for all experiments.
• We tune the droppath of Swin Large backbone from 0.3 to 0.5 for experiments of baseline and our method.

Installation

We test our models under python=3.7.10,pytorch=1.10.1,cuda=10.2. Other versions might be available as well.

Please follow get_started.md to install the repo.

Run

To train a model using 8 cards

bash ./tools/dist_train.sh <config_path> 8

To eval a model using 8 cards

bash ./tools/dist_test.sh  <config_path> <checkpoint_path> 8 --eval keypoints

Modified files compared to Opera

To support Hybrid-branch

• opera/models/dense_heads/petr_head.py
• opera/models/dense_heads/__init__.py

To support checkpoint

• opera/models/utils/transformer.py
• opera/models/utils/__init__.py