Skip to content

Latest commit

 

History

History
83 lines (68 loc) · 3.22 KB

README.md

File metadata and controls

83 lines (68 loc) · 3.22 KB
Raw

PI-Net: Pose Interacting Network for Multi-Person Monocular 3D Pose Estimation

This is the code for the paper Wen Guo, Enric Corona, Francesc Moreno-Noguer, Xavier Alameda-Pineda, PI-Net: Pose Interacting Network for Multi-Person Monocular 3D Pose Estimation, in WACV2021.

Dependencies

Our code is tested on CUDA9, Python3.6, Pytorch1.3.0 MATLAB is needed for evaluating the 3DPCK errors.

Directory

ROOT
|-- data
|-- model
|-- utils 
`-- output
    |-- log
    |-- result 
    |-- snapshot 
        `-- snapshot_24.pth.tar
    |-- tensorboard_log 
    `-- vis 

data
|-- MuCo
    |-- MuCo.py
    `-- data
        |-- augmented_set
        |-- annotations
            |-- MuCo-3DHP_with_posenent_result_filter.json
            |-- MuCo_id2pairId.json
            `-- split_gt.py
|--MuPoTS_skeleton
    |-- MuPoTS_skeleton.py
    |-- bbox_root
    `-- data
        |-- MultiPersonTestSet
        |-- eval
        |-- MuPoTS-3D_with_posenent_result.json
        |-- MuPoTS-3D_id2pairId.json 
        `-- split_gt.py

Preparing data

  • Download Training and testing data MuCo and MuPoTS from SingleShot or from 3DMPPE.
  • Run baseline model 3DMPPE to get pririor poses, and save the result in MuCo-3DHP_with_posenent_result_filter.json (To save the result, please refer to evaluation code in data/MuCo/MuCo.py). If you want to work on another baseline, just save the results in the same format.
  • Easy start: MuCo-3DHP_with_posenent_result_filter.json, MuCo_id2pairId.json, MuPoTS-3D_with_posenent_result.json, and our pretrained model snapshot_24.pth.tar could be downloaded here.

Training and testing

  • To run train the model and test on MPJPE, please just uncommand the corresponding line in model/run.sh and run it directly.
  • To evaluate the result by 3DPCK, please use the --save_mat_result option to get the 2D and 3D result in .m, and use the evaluation code in SingleShot. You can follow the instructions in 3DMPPE to download and set up the evaluation codes.

Citing

If you use our code, please cite our work @inproceedings{guo2021pi, title={PI-Net: Pose Interacting Network for Multi-Person Monocular 3D Pose Estimation}, author={Guo, Wen and Corona, Enric and Moreno-Noguer, Francesc and Alameda-Pineda, Xavier}, booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision}, pages={2796--2806}, year={2021} }

Acknowledgments

The overall code framework is adapted from 3DMPPE and Torchseg. The predictor model code is adapted from SeeWoLook.

Licence

MIT