Pytorch Pose Estimation Framework

A pytorch pose estimation framework using by myself for research.
It contains a pytorch framework which is suitable for pose estimation.
The main partion of it focus on OpenPose reproduction, with simliar mAP in their paper.
You can also found code for my HSI paper in train_offset and train_mask.

Content

1. Requirement
2. Test
3. Train
4. Result
5. Citation
6. License

Requirement

1. Pytorch 1.2.0
2. Torchvision 0.4.0
3. TensorboardX This repro using the environment created by Anaconda, cudatoolkits and cudnn installed by conda automaticly.
   

Test

1. for test, download our pretrained model in dropbox, you also need parpare data by bash get_data.sh
2. open evaluate.py, all parameters related to evaluation are shown here.
   1160 means OpenPose offical small val dataset, which has been uesd in OpenPose paper.
   others means COCO val 2017, which contains 2693 images or 5000 images, choosen by yourself.
   scale should be [0.5,1.0,1.5,2.0] to get the maximum accuracy, as same as OpenPose
3. run python -m Pytorch_Pose_Estimation_Framework.evaluate --val_type=1160 --network=CMU_old --scale=0.5,1.0,1.5,2.0 to get the mAP result
4. if you want to run yourselves images, just modify the main function in evaluate.py, actually, just some path need to be changed

Train

1. all parameters related to train is in train_op_baseline.py
2. you need prepare the data by bash get_data.sh
3. you need generate mask file by running generate_mask.py
4. you need generate hdf5 file for training, about 200G disk space is needed. run generate_hdf5.py
5. check the data path in datasets/dataloader/cmu_h5_mainloader
6. run python -m Pytorch_Pose_Estimation_Framework.train_op_baseline --network=CMU_old

Result

We empirically trained the model for 55 epochs and achieved comparable performance to the results reported in the original paper.
We also compared with the offical released caffe model which is by Zhe Cao.

Method	Validation	AP
Openpose paper	COCO2014-Val-1k	58.4
Openpose model	COCO2014-Val-1k	56.3
This repo	COCO2014-Val-1k	58.4

Acknowledgment

This repo is based upon@kevinlin311tw and @tensorboy.
Thanks kevinlin311tw who is really nice to communicate

Citation

Please cite the paper in your publications if it helps your research:

@inproceedings{cao2017realtime,
  author = {Zhe Cao and Tomas Simon and Shih-En Wei and Yaser Sheikh},
  booktitle = {CVPR},
  title = {Realtime Multi-Person 2D Pose Estimation using Part Affinity Fields},
  year = {2017}
  }
  
@inproceedings{liu2020resolution,
    title={Resolution Irrelevant Encoding and Difficulty Balanced Loss Based Network Independent Supervision for Multi-Person Pose Estimation},
    author={Liu, Haiyang and Luo, Dingli and Du, Songlin and Ikenaga, Takeshi},
    booktitle={2020 13th International Conference on Human System Interaction (HSI)},
    pages={112--117},
    year={2020},
    organization={IEEE}
  }