Fabric Material Recovery from Video Using Multi-Scale Geometric Auto-Encoder

Junbang Liang, Ming C. Lin. ECCV 2022.

Project Page

Requirements

• Python 3.6.4
• PyTorch tested on version 1.1.0
• trimesh
• opencv-python
• smplx
• chumpy

Train

• If you wish to train the garment geometry auto-encoder:

python train.py --train_module atlas

• If you wish to train the single-frame body and garment estimation module:

python train.py --train_module single

• If you wish to train the garment material estimation module:

python train.py --train_module material

Data

Training/test dataset, as well as pretrained model, can be found at here. Please unzip them and place the four folders under the folder of this repo.

Demo

• If you wish to test the single-frame reconstruction only:

python realworld_demo.py --img ${your_test_image_path}

• If you wish to test the material estimation:

python video_demo.py --img ${path_of_splitted_video_frame_images}

Citation

If you use this code for your research, please consider citing:

@inProceedings{liang2022fabric,
  title={Fabric Material Recovery from Video Using Multi-Scale Geometric Auto-Encoder},
  author = {Junbang Liang and Ming C. Lin},
  booktitle={ECCV},
  year={2022}
}

Acknowledgement

This code repository is based on SPIN.