CVPR 2024
@inproceedings{wen2024gomavatar,
title={{GoMAvatar: Efficient Animatable Human Modeling from Monocular Video Using Gaussians-on-Mesh}},
author={Jing Wen and Xiaoming Zhao and Zhongzheng Ren and Alex Schwing and Shenlong Wang},
booktitle={CVPR},
year={2024}
}Our codes are tested in
- CUDA 11.6
- PyTorch 1.13.0
- PyTorch3D 0.7.0
Install the required packages:
conda create -n GoMAvatar
conda activate GoMAvatar
conda install pytorch==1.13.0 torchvision==0.14.0 pytorch-cuda=11.6 -c pytorch -c nvidia
pip install -r requirements.txt
# install pytorch3d
conda install -c fvcore -c iopath -c conda-forge fvcore iopath
conda install pytorch3d -c pytorch3d
# install gaussian splatting
pip install git+"https://github.com/graphdeco-inria/diff-gaussian-rasterization"Download SMPL v1.0.0's male/female models from here and SMPL's neutral model from here and put the .pkl files under utils/smpl/models.
You may need to remove the Chumpy objects following here.
First download the ZJU-MoCap dataset and save the raw data under data/zju-mocap.
Run the following script to preprocess the dataset:
cd scripts/prepare_zju-mocap
python prepare_dataset.py --cfg "$SCENE".yamlChange $SCENE to one of 377, 386, 387, 392, 393, 394.
The folder will be in the following structure:
├── data
├── zju-mocap
├── 377
├── 386
├── ...
├── CoreView_377
├── CoreView_386
├── ...Folders named after scene ID only are preprocessed training data, while those prefixed with CoreView_ are raw data.
Download the PeopleSnapshot dataset and save the files under data/snapshot.
Download the refined training poses from here.
Run the following script to preprocess the training and test set.
cd scripts/prepare_snapshot
python prepare_dataset.py --cfg "$SCENE".yaml # training set
python prepare_dataset.py --cfg "$SCENE"_test.yaml # test set$SCENE is one of female-3-casual, female-4-casual, male-3-casual and male-4-casual.
After the preprocessing, the folder will be in the following structure:
├── data
├── snapshot
├── f3c_train
├── f3c_test
├── f4c_train
├── f4c_test
├── ...
├── female-3-casual
├── female-4-casual
├── ...
├── poses # refined training poses
├── female-3-casual
├── poses
├── anim_nerf_test.npz
├── anim_nerf_train.npz
├── anim_nerf_val.npz
├── ...
Folders ended with _train or _test are preprocessed data.
We provide the pretrained checkpoints in this link. To reproduce the rendering results in the paper, run
# ZJU-MoCap novel view synthesis
python eval.py --cfg exps/zju-mocap_"$SCENE".yaml --type view
# ZJU-MoCap novel pose synthesis
python eval.py --cfg exps/zju-mocap_"$SCENE".yaml --type poseOn the PeopleSnapshot dataset, we follow Anim-NeRF and InstantAvatar to refine test poses:
python train_pose.py --cfg exps/snapshot_"$SCENE".yamlPlease check exps/ for detailed configuration files.
You can run 360 degree freeview rendering using the following command
python eval.py --cfg exps/zju-mocap_"$SCENE".yaml --type freeviewUse --frame_idx to specify the training frame id and --n_frames to set the number of views.
Or you can render novel poses from MDM:
python eval.py --cfg exps/zju-mocap_"$SCENE".yaml --type pose_mdm --pose_path data/mdm_poses/sample.npyWe provide an example of pose trajectory in data/pose_mdm/sample.npy.
Run the following command to train from scratch:
# ZJU-MoCap
python train.py --cfg exps/zju-mocap_"$SCENE".yaml
# PeopleSnapshot
python train.py --cfg exps/snapshot_"$SCENE".yamlThis project builds upon HumanNeRF and MonoHuman. We appreciate the authors for their great work!