GoMAvatar: Efficient Animatable Human Modeling from Monocular Video Using Gaussians-on-Mesh

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}
}

Requirements

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"

Data preparation

Prerequisites

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.

ZJU-MoCap

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".yaml

Change $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.

PeopleSnapshot

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.

Rendering and evaluation

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 pose

On the PeopleSnapshot dataset, we follow Anim-NeRF and InstantAvatar to refine test poses:

python train_pose.py --cfg exps/snapshot_"$SCENE".yaml

Please 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 freeview

Use --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.npy

We provide an example of pose trajectory in data/pose_mdm/sample.npy.

Training

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".yaml

Acknowledgements

This project builds upon HumanNeRF and MonoHuman. We appreciate the authors for their great work!

wenj / GoMAvatar

readme