A Lightweight Graph Transformer Network for Human Mesh Reconstruction from 2D Human Pose
The project is an official implementation of our paper A Lightweight Graph Transformer Network for Human Mesh Reconstruction from 2D Human Pose.
Installation
Check INSTALL.md for installation instructions.
Quick demo
We provide demo codes to run end-to-end inference on the test images. Please check DEMO.md for details.
Datasets Download
Please download the required datasets following DOWNLOAD.md.
Experiment
The experiment
directory should contain following folders.
${ROOT}
|-- experiment
| |-- pam_h36m
| | |-- best.pth.tar
| |-- gtrs_h36m
| | |-- final.pth.tar
| |-- pam_3dpw
| | |-- best.pth.tar
| |-- pam_3dpw
| | |-- final.pth.tar
Pretrained model weights
The pretrained model weights can be download from here to a corresponding directory.
Testing
You can choose the config file in ${ROOT}/asset/yaml/
to evaluate the corresponding experiment by running:
python main/test.py --gpu 0,1, --cfg ./asset/yaml/gtrs_{input joint set}_test_{dataset name}.yml
For example, if you want to test the results on Human3.6M dataset, you can run:
python main/test.py --gpu 0,1, --cfg ./asset/yaml/gtrs_human36J_test_human36.yml
Training
We provide the pretrain models of PAM module. To train GTRS, you can run:
python main/train.py --gpu 0,1, --cfg ./asset/yaml/gtrs_{input joint set}_train_{dataset name}.yml
For example, if you want to train on Human3.6M dataset, you can run:
python main/train.py --gpu 0,1, --cfg ./asset/yaml/gtrs_human36J_train_human36.yml
Also if you prefer training from the scratch, you should pre-train PAM module first by running:
python main/train.py --gpu 0,1, --cfg ./asset/yaml/pam_{input joint set}_train_{dataset name}.yml
Citations
If you find our work useful in your research, please consider citing:
@article{ce2021gtrs,
title={A Lightweight Graph Transformer Network for Human Mesh Reconstruction from 2D Human Pose},
author={Zheng, Ce and Mendieta, Matias and Wang, Pu and Lu, Aidong and Chen, Chen},
journal={arXiv preprint arXiv:2111.12696},
year={2021}
}
License
Our research code is released under the MIT license. See LICENSE for details.
Acknowledgments
Our implementation and experiments are built on top of open-source GitHub repositories. We thank all the authors who made their code public, which tremendously accelerates our project progress. If you find these works helpful, please consider citing them as well.