Recruitment from 2012 Labs of Huawei

Who are we? Our team is the CG & 3DV center of Huawei. We are working for bringing the 3D digital world to everyone.
What are we doing? Our main research lies on human-centric 3D/4D reconstruction, animation, simulation and generation.
Who are we looking for? Everyone (mainly students) working on 3D vision and graphics, looking at this repo :)
Homepage of the contact: 李哲. Welcome to chat!
Wechat: nexus_unite, E-mail: lizhe_thu[AT]126.com

News

09/15/2024 Release the templates of ActorsHQ (Actor01 & Actor04) to facilitate training.
05/22/2024 📢 An extension work of Animatable Gaussians for human avatar relighting is available here. Welcome to check it!
03/11/2024 The code has been released. Welcome to have a try!
03/11/2024 AvatarReX dataset, a high-resolution multi-view video dataset for avatar modeling, has been released.
02/27/2024 Animatable Gaussians is accepted by CVPR 2024!

Animatable Gaussians: Learning Pose-dependent Gaussian Maps for High-fidelity Human Avatar Modeling

CVPR 2024

Zhe Li ¹, Zerong Zheng ², Lizhen Wang ¹, Yebin Liu ¹

¹Tsinghua Univserity ²NNKosmos Technology

Projectpage · Paper · Video

teaser.mp4

Abstract: Modeling animatable human avatars from RGB videos is a long-standing and challenging problem. Recent works usually adopt MLP-based neural radiance fields (NeRF) to represent 3D humans, but it remains difficult for pure MLPs to regress pose-dependent garment details. To this end, we introduce Animatable Gaussians, a new avatar representation that leverages powerful 2D CNNs and 3D Gaussian splatting to create high-fidelity avatars. To associate 3D Gaussians with the animatable avatar, we learn a parametric template from the input videos, and then parameterize the template on two front & back canonical Gaussian maps where each pixel represents a 3D Gaussian. The learned template is adaptive to the wearing garments for modeling looser clothes like dresses. Such template-guided 2D parameterization enables us to employ a powerful StyleGAN-based CNN to learn the pose-dependent Gaussian maps for modeling detailed dynamic appearances. Furthermore, we introduce a pose projection strategy for better generalization given novel poses. Overall, our method can create lifelike avatars with dynamic, realistic and generalized appearances. Experiments show that our method outperforms other state-of-the-art approaches.

Demo Results

We show avatars animated by challenging motions from AMASS dataset.

basketball.mp4

More results (click to expand)

football.mp4

dancing.mp4

irish_dancing.mp4

Installation

Clone this repo.

git clone https://github.com/lizhe00/AnimatableGaussians.git
# or
git clone git@github.com:lizhe00/AnimatableGaussians.git

Install environments.

# install requirements
pip install -r requirements.txt

# install diff-gaussian-rasterization-depth-alpha
cd gaussians/diff_gaussian_rasterization_depth_alpha
python setup.py install
cd ../..

# install styleunet
cd network/styleunet
python setup.py install
cd ../..

Download SMPL-X model, and place pkl files to ./smpl_files/smplx.

Data Preparation

AvatarReX, ActorsHQ or THuman4.0 Dataset

Download AvatarReX, ActorsHQ, or THuman4.0 datasets.
Data preprocessing. We provide two manners below. The first way is recommended if you plan to employ our pretrained models, because the renderer utilized in preprocessing may cause slight differences.
1. (Recommended) Download our preprocessed files from PREPROCESSED_DATASET.md, and unzip them to the root path of each character.
2. Follow the instructions in gen_data/GEN_DATA.md to preprocess the dataset.

Note for ActorsHQ dataset: 1) DATA PATH. The subject from ActorsHQ dataset may include more than one sequences, but we only utilize the first sequence, i.e., Sequence1. The root path is ActorsHQ/Actor0*/Sequence1. 2) SMPL-X Registration. We provide SMPL-X fitting for ActorsHQ dataset. You can download it from here, and place smpl_params.npz at the corresponding root path of each subject.

Customized Dataset

Please refer to gen_data/GEN_DATA.md to run on your own data.

Avatar Training

Take avatarrex_zzr from AvatarReX dataset as an example, run:

python main_avatar.py -c configs/avatarrex_zzr/avatar.yaml --mode=train

After training, the checkpoint will be saved in ./results/avatarrex_zzr/avatar.

Avatar Animation

Download pretrained checkpoint from PRETRAINED_MODEL.md, unzip it to ./results/avatarrex_zzr/avatar, or train the network from scratch.
Download THuman4.0_POSE or AMASS dataset for acquiring driving pose sequences. We list some awesome pose sequences from AMASS dataset in configs/awesome_amass_poses.yaml. Specify the testing pose path in configs/avatarrex_zzr/avatar.yaml#L57.
Run:

python main_avatar.py -c configs/avatarrex_zzr/avatar.yaml --mode=test

You will see the animation results like below in ./test_results/avatarrex_zzr/avatar.

example_animation.mp4

Evaluation

We provide evaluation metrics and example codes of comparison with body-only avatars in eval/comparison_body_only_avatars.py.

Todo

Release the code.
Release AvatarReX dataset.
~~Release all the checkpoints and preprocessed dataset.~~ Cancelled due to graduation. Please run on other cases yourself with the provided configs.

Acknowledgement

Our code is based on these wonderful repos:

3D Gaussian Splatting and its adapted version
StyleAvatar

Citation

If you find our code or data is helpful to your research, please consider citing our paper.

@inproceedings{li2024animatablegaussians,
  title={Animatable Gaussians: Learning Pose-dependent Gaussian Maps for High-fidelity Human Avatar Modeling},
  author={Li, Zhe and Zheng, Zerong and Wang, Lizhen and Liu, Yebin},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2024}
}