DisCo: Disentangled Implicit Content and Rhythm Learning for Diverse Co-Speech Gestures Synthesis

Reproduction Based on Relased Model

Train and test DisCo

0. python == 3.7
1. build folders like:

   audio2pose
   ├── codes
   │ └── audio2pose
   ├── datasets
   │ ├── trinity
   │ └── s2g
   └── outputs
   └── audio2pose
   ├── custom
   └── wandb 

2. download the framework scripts from BEAT to codes/audio2pose/
3. run pip install -r requirements.txt in the path ./codes/audio2pose/
4. download trinity dataset to datasets/trinity
5. bulid data cache and calculate mean and std by given number of joints, FPS, speakers using /dataloader/preprocessing.ipynb
6. put disco.py under ./audio2pose/model/ and customize disco_trainer.py for contrastive learning.
7. run python train.py -c ./configs/disco_trinity_ae.yaml for pretrained_ae for FID calculation.
8. run python train.py -c ./configs/disco_trinity.yaml for training.
9. run python test.py -c ./configs/disco_trinity.yaml for inference.
10. load ./outputs/audio2pose/custom/exp_name/epoch_number/xxx.bvh into blender to visualize the test results.

From data in other dataset (e.g. S2G)

• refer train and test DisCo for bvh cache
• set dataset: trinity in .yaml

Citation

DisCo is established for the following research project.

@inproceedings{liu2022disco,
title={DisCo: Disentangled Implicit Content and Rhythm Learning for Diverse Co-Speech Gestures Synthesis},
author={Liu, Haiyang and Iwamoto, Naoya and Zhu, Zihao and Li, Zhengqing and Zhou, You and Bozkurt, Elif and Zheng, Bo},
booktitle={Proceedings of the 30th ACM International Conference on Multimedia},
pages={3764--3773},
year={2022}
}