Pytorch Implementation of Robust Motion In-betweening

This is the unofficial implementation of the approach described in the paper:

Felix G. Harvey, Mike Yurick, Derek Nowrouzezahrai, and Christopher Pal Robust Motion In-betweening. In ACM Transactions on Graphics (TOG), 2020.

We provide the code for reproducing the main results, as well as pre-trained models.

Dependencies

• Python 3+ distribution
• PyTorch >= 1.4.0
• NumPy
• PIL
• TensorboardX
• Pyyaml

Please follow this repo to download the data. Pretrained model is available at this link. After downloading this repo, you need: (1) create new dirs named src, log, model, gif, and results repectively; (2) Put all downloaded files in to ./src and pretrained model into ./model.

For data preparation

python flip_bvh.py

For training

python train.py

For testing

python test.py

The contribution of foot sliding loss

In the original papaer, foot sliding problem is only post processed. Here I add the foot sliding loss which turned out to be effective to further enhance visual quality. Here is an exmaple: Images from left to right are orginal implementation, + foot sliding loss, + IK post processing, and ground truth respecitvely. With the help of foot sliding loss, the model is able to infer a rational foot contact arrange to reach the target.

Work status

This sheet

Demo results

Synthesized resutls without foot sliding constraint could be downloaded from this link, the results with foot sliding constraint could be downloaded from this link.