Unsupervised Shape and Pose Disentanglement for 3D Meshes

Repo for "Unsupervised Shape and Pose Disentanglement for 3D Meshes, ECCV'20 (Poster)"

Link to paper: https://arxiv.org/abs/2007.11341

Link to project: https://virtualhumans.mpi-inf.mpg.de/unsup_shape_pose/

Prerequisites

1. Cuda 9.0
2. Python 2.7
3. Pytorch 1.3
4. Scikit-sparse
5. MPI mesh library (https://github.com/MPI-IS/mesh)
6. OpenDR (https://github.com/mattloper/opendr)

For spiral convolution we use code from Neural3DMM repo and modify it according to our needs.

Data Preprocessing

1. Download and uncompress AMASS Dataset (https://amass.is.tue.mpg.de/)
2. Download SMPL+H body models (https://mano.is.tue.mpg.de/)
3. Preprocess AMASS to generate training/validation/test sets: python data/data_extraction.py

Model Training

1. Edit config.json to use your own directory structures and model hyperparameters
2. Run python train.py

Pretrained Models

You can download pretrained model for AMASS at https://drive.google.com/file/d/1Uge1PKQoL1xy8UH4iLXGz9k-div8aEgu/view?usp=sharing

Please consider citing our work if you found it useful:

@inproceedings{zhou20unsupervised,
    title = {Unsupervised Shape and Pose Disentanglement for 3D Meshes},
    author = {Zhou, Keyang and Bhatnagar, Bharat Lal and Pons-Moll, Gerard},
    booktitle = {European Conference on Computer Vision (ECCV)},
    month = {August},
    year = {2020},
}