xljh0520 / JOTR

Official implementation of “JOTR: 3D Joint Contrastive Learning with Transformers for Occluded Human Mesh Recovery“

JOTR: 3D Joint Contrastive Learning with Transformers for Occluded Human Mesh Recovery

ICCV 2023

Official PyTorch implementation of the paper "JOTR: 3D Joint Contrastive Learning with Transformers for Occluded Human Mesh Recover"

[Preprint]

Installation

conda create python=3.7 --name jotr
conda activate jotr
pip install torch==1.8.0 torchvision==0.9.0
sh requirements.sh

Data Preparation

We prepare the data in a samilar way as 3DCrowdNet. Please refer to 3DCrowdNet for dataset, SMPL model, VPoser model, and backbone pre-trained weights.

Download the annotations of 3DPW-PC, 3DPW-OC.

Download checkpoint of JOTR from here.

The data directory should be organized as follows:

${ROOT}  
|-- checkpoint
|-- 3dpw_best_ckpt.pth.tar
|-- 3dpw-crowd_best_ckpt.pth.tar
|-- 3dpw-oc_best_ckpt.pth.tar
|-- 3dpw-pc_best_ckpt.pth.tar
|-- data 
|   |-- J_regressor_extra.npy 
|   |-- snapshot_0.pth.tar
|   |-- 3DPW
|   |   |-- 3DPW_latest_test.json
|   |   |-- 3DPW_oc.json
|   |   |-- 3DPW_pc.json
|   |   |-- 3DPW_validation_crowd_hhrnet_result.json
|   |   |-- imageFiles
|   |   |-- sequenceFiles
|   |-- CrowdPose
|   |   |-- annotations
|   |   |-- images
|   |-- Human36M  
|   |   |-- images  
|   |   |-- annotations   
|   |   |-- J_regressor_h36m_correct.npy
|   |-- MSCOCO  
|   |   |-- images  
|   |   |   |-- train2017  
|   |   |-- annotations  
|   |   |-- J_regressor_coco_hip_smpl.npy
|   |-- MuCo  
|   |   |-- augmented_set  
|   |   |-- unaugmented_set  
|   |   |-- MuCo-3DHP.json
|   |   |-- smpl_param.json

Evaluation

Reproduce the results in the paper (Table 1 and Table 2) by running the following command:

sh eval.sh

Training

Train the model by running the following command:

sh train.sh

Visualization

TODO

Acknowledgments

Thanks to 3DCrowdNet, DETR, AutomaticWeightedLoss, deep_training and PositionalEncoding2D, our code is partially borrowing from them.

License

This code is distributed under an MIT LICENSE.

Note that our code depends on other libraries, including SMPL, VPoser, and uses datasets that have their own licenses. Please refer to the corresponding websites for more details.

Citation

@article{li2023jotr,
  title={JOTR: 3D Joint Contrastive Learning with Transformers for Occluded Human Mesh Recovery},
  author={Li, Jiahao and Yang, Zongxin and Wang, Xiaohan and Ma, Jianxin and Zhou, Chang and Yang, Yi},
  journal={ICCV},
  year={2023}
}