Official Repository for the newly published paper entitled "HMC: Hierarchical Mesh Coarsening for Skeleton-free Motion Retargeting"
[Project Page] [ArXiv Page] [Paper]
conda install --yes --file requirements.txt
pip install -r requirements.txt
1. download the pretrained model & demo data from Google Drive
Drop the pretrained model into pretrained/
and demo data into data/
, respectively.
Here are backup links to Baidu Disk for both model & data.
Run
python inference_hmc.py
Then, a motion sequence as greeting_on_target-XXXXXX.obj
will be saved in data\greeting_on_target\
.
If the retargeted sequence is converted to .abc
format (a routine for automatic conversion will be provided in the future), it should be like this:
Create a source folder data/{src_name}
comprising source T-pose (data/{src_name}/{src_name}-tpose.obj
) and motion sequence (data/{src_name}/{src_name}-{idx}.obj
, where {idx}
is counted from
Create a target folder data/{tgt_name}
comprising only target T-pose (data/{tgt_name}/{tgt_name}-tpose.obj
).
Run
python inference_hmc.py --src_name={src_name} --tgt_name={tgt_name}
and a motion sequence of the same length as source motion will be produced in folder data/{tgt_name}/
As an alternative to 3, run
python inference_hmc.py --src_name={src_name} --tgt_name={tgt_name} --precoarsen_src={pc_ratio}
where {pc_ratio}
is a continuous value in {pc_ratio}
, the retargeting process can be accelerated, and the model also considers few mesh details on the source. However, one should note that in some cases, a small {pc_ratio}
may induce extra jitters in target motion due to uncaught local motions.
If you use HMC in any context, please cite the following paper:
@misc{wang2023hmc,
title={HMC: Hierarchical Mesh Coarsening for Skeleton-free Motion Retargeting},
author={Haoyu Wang and Shaoli Huang and Fang Zhao and Chun Yuan and Ying Shan},
year={2023},
eprint={2303.10941},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
}