This repository includes the experiment code and trained model of paper "Disentangled Human Body Embedding Based on Deep Hierarchical Neural Network", TVCG 2020, https://arxiv.org/abs/1905.05622. Authors: Boyi Jiang, Juyong Zhang, Jianfei Cai and Jianmin Zheng
The transformed connectivity consistent human body dataset, including the ACAP feature and mesh file, can be downloaded here. Note that these meshes are transformed from partial data of other datasets, including SCAPE, FAUST, Dyna, MANO, Hasler et al and SPRING. No commercial usage of the data is allowed.
- Ubuntu, TITAN Xp, CUDA-10.0, Python 3.7
- Python packages: PyTorch(1.3.0), PyTorch geometric(1.3.2), Numpy, openmesh, opencv, batch_knn(some experiment codes need).
- Make a folder called "models" in the repository directory.
- Download the trained model and unzip to the "models" folder..
- Download the trained extension model and unzip to the "models" folder.(Optional)
The "exps" folder includes sample codes with test data of several experiments described in the paper.
- PoseTransfer：A sample code of pose transfer, running this code should produce the results of figure 13 in the paper.
- point2point: A sample code for registering meshes with same connectivity.
- DFaustFit: A sample code of the DFaust registration experiment. (about 10 minutes)
- FaustAlign: A sample code of Faust alignment experiment.(about 20 minutes)
- BodyFrom2DJoints: A sample code of 3D body pose estimation from 2D joints.
Please cite the following paper if it helps your research:
@article{Jiang2020HumanBody,
title={Disentangled Human Body Embedding Based on Deep Hierarchical Neural Network},
author={Jiang, Boyi and Zhang, Juyong and Cai, Jianfei and Zheng, Jianmin},
booktitle={IEEE Transactions on Visualization and Computer Graphics},
year={2020}
}