cjm-sfw/querypose

This is an official implementation of our NeurIPS 22 paper“QueryPose: Sparse Multi-Person Pose Regression via Spatial-Aware Part-Level Query”

QueryPose: Sparse Multi-Person Pose Regression via Spatial-Aware Part-Level Query

👏👏👏👏👏👏👏👏👏👏👏 QueryPose is a sparse end-to-end multi-person pose regression framework:

QueryPose: Sparse Multi-Person Pose Regression via Spatial-Aware Part-Level Query,
Yabo Xiao, Kai Su, Xiaojuan Wang, Dongdong Yu, Lei Jin, Mingshu He, Zehuan Yuan;
Published on NeurIPS 2022

Highlights

• Simple: QueryPose is sparse end-to-end multi-person pose regression framework, which can directly predict multi-person keypoint sequences from the input image. Our method is built upon Sparse RCNN without transformer encoder.

• Generalizability: QueryPose is able to achieve the competitive performance on diverise scenes.

• Fast convergence: Achieving 60+ AP with only several training epochs. We will release the 1x model with competitive performance.

Models

The pretrain model of HRNet-series can be downloaded from HRNet-pretrain. The swin-series can be downloaded from Swin Transformer

The results on MS COCO mini-val with 100 queries; We provide the light version models (without dynamic head in keypoint decoder)

Backbone	keypoint AP	keypoint AP *	Times(ms)	download
HRNet32_100pro_3x	69.8	72.3	97	model
HRNet48_100pro_3x	71.0	73.4	101	model

Notes

• We are restructuring our code for accelerating training and inference (reducing the GPU memory significantly as well). Working in Progress 👷‍♂️👷‍♂️👷‍♂️!!! The more effecient implementation will coming soon.
• The time is calculated on Tesla A100, Other GPU cards are not fully tested.
• For QueryPose, we leverage the heatmap-pretrain model for HRNet-series backbone instead of using online auxiliary heatmap learning in other regression-based methods. We observe that the online auxiliary heatmap learning is useless for our method.

Installation

The codebases are built on top of Detectron2 and Sparse RCNN. If you have any environment or compilation issues, please refer to them.

1. Install and compile

git clone [email protected]:buptxyb666/QueryPose.git
cd QueryPose
python setup.py build develop

pip install timm

2. Link coco dataset path to QueryPose/datasets/coco

mkdir -p datasets/coco
ln -s /path_to_coco_dataset/annotations datasets/coco/annotations
ln -s /path_to_coco_dataset/train2017 datasets/coco/train2017
ln -s /path_to_coco_dataset/val2017 datasets/coco/val2017

3. Train QueryPose

python projects/querypose/train_net.py --num-gpus 8 \
    --config-file projects/querypose/configs/querypose.hrnet32.100pro.3x.yaml

4. Evaluate QueryPose

python projects/querypose/train_net.py --num-gpus 8 \
    --config-file projects/querypose/configs/querypose.hrnet32.100pro.3x.yaml \
    --eval-only MODEL.WEIGHTS path/to/model.pth

Citing

If you find this project useful for your research, please use the following BibTeX entry:

@inproceedings{xiaoquerypose,
  title={QueryPose: Sparse Multi-Person Pose Regression via Spatial-Aware Part-Level Query},
  author={Xiao, Yabo and Su, Kai and Wang, Xiaojuan and Yu, Dongdong and Jin, Lei and He, Mingshu and Yuan, Zehuan},
  booktitle={Advances in Neural Information Processing Systems}
}