abcyzj/Semantics2Hands

Official implementation for the paper "Semantics2Hands: Transferring Hand Motion Semantics between Avatars".

Semantics2Hands

Official implementation for the paper "Semantics2Hands: Transferring Hand Motion Semantics between Avatars".

Project Page | Paper

Bibtex

@inproceedings{ye2023semantics,
  title={Semantics2Hands: Transferring Hand Motion Semantics between Avatars},
  author={Ye, Zijie and Jia, Jia and Xing, Junliang},
  booktitle={Proceedings of the 31st ACM International Conference on Multimedia},
  year={2023}
}

Getting Started

1. Create Conda Environment

conda create -n s2h python=3.8
conda activate s2h

The code was tested on Python 3.8, PyTorch 1.13.1 and CUDA 12.1.

2. Install Dependencies

Install packages from requirements.txt.

pip install -r requirements.txt

Install PyTorch3D. Following the official instruction of PyTorch3D, you can install with

conda install -c fvcore -c iopath -c conda-forge fvcore iopath
pip install --no-index --no-cache-dir pytorch3d -f https://dl.fbaipublicfiles.com/pytorch3d/packaging/wheels/py38_cu113_pyt1121/download.html # Installing from prebuilt wheels is faster than using conda, but the prebuilt wheel url may differ depeding on your CUDA version.

The chumpy package required by smplx is no longer maintained and would raise ImportError. We need to comment line 11 in "CONDA_DIR/envs/s2h/lib/python3.8/site-packages/chumpy/init.py". CONDA_DIR is the directory where you install conda.

from numpy import bool, int, float, complex, object, unicode, str, nan, inf
# comment the above line in "CONDA_DIR/envs/s2h/lib/python3.8/site-packages/chumpy/__init__.py", line 11

3. Install Blender

Install Blender >= 2.82 from: https://www.blender.org/download/.

4. Download the MANO hand model

Download the MANO model file from MANO. Place MANO_LEFT.pkl and MANO_RIGHT.pkl in artifact/smplx/models/mano.

5. Data Preparation

We use the Mixamo and InterHand2.6M datasets to train our model. You can download the preprocessed data from Google Drive. Then place the MixHand directory in artifact directory.

Otherwise, if you want to preprocess the dataset on your own or use different characters, please follow the instructions in DATA.md.

Using the pretrained model

First, download the pretrained model and place it at artifact.

1. Inter-Domain Retargeting

To reproduce quantitative results, run:

python -m run.train_mixhand test --config artifact/ASRN/lightning_logs/version_0/config.yaml --ckpt_path artifact/ASRN/lightning_logs/version_0/checkpoints/best_mixamo-epoch=58-mixamo_semi_rf=-1.80.ckpt

To reproduce qualitative results, run:

python -m run.visualize_mixhand --ckpt_path artifact/ASRN/lightning_logs/version_0/checkpoints/best_mixamo-epoch=58-mixamo_semi_rf=-1.80.ckpt --config ./artifact/ASRN/lightning_logs/version_0/config.yaml --output_dir artifact/visualization

2. Intro-Domain Retargeting

First, uncomment line 180~183 in data/combined_motion.py.

To reproduce quantitative results, run:

python -m run.train_mixhand test --config artifact/ASRN/lightning_logs/version_0/config.yaml --ckpt_path artifact/ASRN/lightning_logs/version_0/checkpoints/best_mixamo-epoch=58-mixamo_semi_rf=-1.80.ckpt

Train from scratch

Run:

python -m run.train_mixhand fit --config config/ASRN.yaml --trainer.default_root_dir artifact/ASRN

The DM baseline

Training

Run:

python -m run.train_dm fit --config config/DM.yaml --trainer.default_root_dir artifact/DM

Inter-Domain Retargeting

To reproduce quantitative results, run:

python -m run.train_dm test --config artifact/DM/lightning_logs/version_0/config.yaml --ckpt_path artifact/DM/lightning_logs/version_0/checkpoints/best_mixamo-epoch=93-mixamo_semi_rf=0.00.ckpt

The name of the checkpoint file may differ.

Intro-Domain Retargeting

First, uncomment line 180~183 in data/combined_motion.py.

To reproduce quantitative results, run:

python -m run.train_dm test --config artifact/DM/lightning_logs/version_0/config.yaml --ckpt_path artifact/DM/lightning_logs/version_0/checkpoints/best_mixamo-epoch=93-mixamo_semi_rf=0.00.ckpt

The name of the checkpoint file may differ.

Acknowledgements

The code of the dataloader, the BVH parser and the Animation object are based on SAN repository.

License

This code is distributed under an MIT LICENSE.