PoseDet: Towards Real-time Multi-person Pose Estimation

Official PyTorch implementation for PoseDet: Fast Multi-Person Pose Estimation Using Pose Embedding. A simple system that performs real-time multi-person pose estimation.

Installation

This repository is based on mmdetection2.1.0. To run PoseDet, you are supposed to install mmdetection first.

Requirements

• Linux
• Python 3.6+, PyTorch 1.3.1, CUDA 9.2 on V100 or GTX 1080 Ti （Pytorch 1.7.1, CUDA 11.1, on RTX 3090 is also supported, other versions are not tested）
• GCC 5+

# A script with Anaconda
cd ./PoseDet_mmdetection
# Create env of conda
conda create -n PoseDet python=3.7
conda activate PoseDet

# Install pytorch
conda install pytorch=1.3.1 cudatoolkit=9.2 torchvision=0.4.2 -c pytorch

# Install mmcv, coco api, and other packages in requirements.
pip install mmcv==0.6.2 --user
pip install -r requirements/build.txt
#TO isntall coco api, you can also download the package and install it localy via python setup.py develop
pip install "git+https://github.com/open-mmlab/cocoapi.git#subdirectory=pycocotools" --user

# Install mmdet by develop mode
python setup.py develop

Quick Start

For model training

You are supposed to download the imagenet pre-trained models, then put them into PoseDet_mmdetection/pretrained:

• DLA 34
• HRNet-W32
• HRNer-W48

Dataset preparation

Download the required dataset and modify the path to the dataset (data_root) in PoseDet/configs/_base_/dataset_coco.py

Train

#Replace config_file with the path to the config file (e.g., ./PoseDet/config/COCO/PoseDet_DLA34_coco.py)
#Config image_per_gpu in terms of the memory limitation, batch size = image_per_gpu * nproc_per_node. The default batch size 32 is most likely the best.
python -m torch.distributed.launch --nproc_per_node=8  tools/train.py --launcher pytorch --config config_file --image_per_gpu 4

Test

#Replace config_file and checkpoint_file
python tools/test.py --config config_file --checkpoint checkpoint_file

#Visulize
#The output image with predicted pose will be saved in show_dir
 python tools/test.py --config config_file --checkpoint checkpoint_file --show-pose --show-dir show_dir

For multi-scale testing and flip, you are supposed to modify the config file (cfg.test_pipeline.MultiScaleFlipAug)

Citation

If you find our work helpful for your research, please consider to cite:

@article{tian2021posedet,
  title={PoseDet: Fast Multi-Person Pose Estimation Using Pose Embedding},
  author={Tian, Chenyu and Yu, Ran, and Zhao, Xinyuan, and Xia, Weihao and Wang, Haoqian and Yang, Yujiu},
  journal={arxiv preprint arxiv:2107.10466},
  year={2021}
}