Skeleton-free Pose Transfer for Stylized 3D Characters

teaser

This is the official repository for ECCV 2022 paper Skeleton-free Pose Transfer for Stylized 3D Characters.

More detailed documentation coming soon!

Prerequisites

Python >= 3.7
Pytorch >= 1.4
Pytorch-geometric
Pymesh (optional)

pip install opencv-python tensorboardx smplx pyrender open3d cython kornia

Demo

Download our demo data and pretrained model from here. Unzip it to the project root directory.

Then,

python demo.py

Checkt the results in ./demo/results and they should be the same as meshes in ./demo/results_reference.

To try with your own data, make sure the number of triangles is around 5K (not a strict requirement) and the orientation of the character is the same as demo data (front: +Z, up: +Y)

Demo with SMPL

Download SMPLH model from here and extract. Set the SMPLH_PATH in global_var.py to the folder it is extracted to.

Then, run

python demo_smpl.py

Training

More documentations about training will come soon.

Data Preparation

All paths are defined in global_var.py

SMPL

Download SMPLH model from here and extract. Set the SMPLH_PATH in global_var.py to the folder it is extracted to. Here is the simplified SMPL mesh used in our training.

Mixamo

The training/testing split we used for the paper can be downloaded here.

Download Mixamo characters of T-pose.
Download Mixamo animations. Note that we don't use character-specific animations. Thus, simply download animations for one character.
Convert static characters (T-pose) into .obj format and animations into .bvh format. Unfortunately, I used some internal codes for this part, which cannot be open-sourced. You might have to write the conversion yourself.
Put characters of T-pose in {MIXAMO_PATH}/obj in .obj format.
Extract skinning information running data_proc/mixamo_fbx_parser.py in Maya, and put ground truth skinning information of characters in {MIXAMO_PATH}/rig_info. The data format should be the same as in rig_info in RigNet dataset.
Run python data_proc/mixamo_simplify.py to make sure the triangle numbers are less than 5000.
Run python data_proc/mixamo_preproc_simplify.py
Run python data_proc/mixamo_motion_preproc.py

Run Training

python train.py

Citation

Please cite our paper if you use this repository:

@inproceedings{liao2022pose,
    title = {Skeleton-free Pose Transfer for Stylized 3D Characters},
    author = {Liao, Zhouyingcheng and Yang, Jimei and Saito, Jun and Pons-Moll, Gerard and Zhou, Yang},
    booktitle = {European Conference on Computer Vision ({ECCV})},
    month = {October},
    organization = {{Springer}},
    year = {2022},
}

Credit

We borrowed part of the codes from the following projects:

https://github.com/zycliao/TailorNet_dataset
https://github.com/zhan-xu/RigNet
https://github.com/YadiraF/face3d
https://github.com/kzhou23/shape_pose_disent

zycliao / skeleton-free-pose-transfer

readme