Mani-GS: Gaussian Splatting Manipulation with Triangular Mesh

Xiangjun Gao¹, Xiaoyu Li², Yiyu Zhuang³, Qi Zhang², Wenbo Hu², Chaopeng Zhang² ,
Yao Yao³ , Ying Shan² Long Quan<sup>1

1</sup>HKUST, ²Tencent, ³Nanjing University

🤗 This is the official implementation for the paper Mani-GS: Gaussian Splatting Manipulation with Triangular Mesh.

🤗 TL;DR: We introduce a Gaussian-Mesh binding strategy with self-adaption for 3DGS manipulation, which can maintain high-quality rendering, have high tolerance for mesh accuracy and support various types of 3DGS manipulations.

📣 News

• [24-6-30] 🔥 Training and inference Code is released.
• [24-5-29] 🔥 Mani-GS is released on arXiv.

🌟 Overview

• 🛠️ Installation
• 📦 Data preparation
• 🚀 Training and Evaluation
• 💫 Manipulation
• 👍 Acknowledgement
• 📘 Citation

🛠️ Installation

Clone this repo

https://github.com/gaoxiangjun/Mani-GS.git

Install dependencies

# install environment
conda env create --file environment.yml
conda activate mani-gs

# install pytorch=1.12.1 and others
conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.6 -c pytorch -c conda-forge
pip install torch_scatter==2.1.1
pip install kornia==0.6.12
pip install imageio[ffmpeg]

# install nvdiffrast=0.3.1
git clone https://github.com/NVlabs/nvdiffrast
pip install ./nvdiffrast

# install knn-cuda
git clone https://github.com/unlimblue/KNN_CUDA.git
cd KNN_CUDA
make && make install

# install relightable 3D Gaussian
pip install ./r3dg-rasterization

📦 Data preparation

NeRF Synthetic Dataset

Download the NeRF synthetic dataset from LINK provided by NeRF.

DTU Dataset

For real-world DTU data, we adopt the Relightable3DGaussian pre-processed DTU data, which can be downloaded here.

For evaluation, we use the NeuMesh test split, which can be download from here, and should be put in datasets folder.

We organize the Data Structure like this:

Mani-GS
├── datasets
    ├── nerf_synthetic
    |   ├── chair
    |   ├── ...
    ├── data_dtu
    |   ├── DTU_scan24
    |   ├── ...
    ├── data_dtu_valnames
    |   ├── dtu_scan24

🚀 Training and Evaluation

The training is divided into two stages: (1) extracting the mesh from 3DGS using Screened Poisson reconstruction or NeuS; and (2) binding 3D Gaussian to a given triangular mesh.

Stage 1 is optional, as we will provide the pre-extracted mesh using different methods. The evaluation will be conducted once the Stage 2 training is complete. Please note that this evaluation is only for static evaluation.

We provide our extracted mesh using different methods, which can be downloaded from this Google drive link. Please unzip it into ./output

NeRF Synthetic dataset:

sh script/run_nerf_stage_1.sh # (optional)
sh script/run_nerf_stage_2.sh # neus mesh as default

DTU data:

sh script/run_dtu_stage_1.sh # (optional)
sh script/run_dtu_stage_2.sh # poisson mesh as default

💫 Manipulation

Data Structure

We also provide a manipuated mesh demo lego in the aforementioned Google Drive link. The provided mesh files are organizea like this:

Mani-GS
├── output
    ├── NeRF_Syn
    |   ├── lego
    |   |   |── mesh_neus_decimate.ply
    |   |   |── mesh_poi_clean.ply
    |   |   |── mesh_mc_30K.ply
    |   |   |── mesh_35K.ply
    |   |   |── final_our_deform_stretch
    |   |   |   |── 00_1.obj
    |   |   |   |── ...
    |   |   |   |── 00_20.obj
    |   |   |── final_our_softbody_35K
    |   |   |   |── 00_1.obj
    |   |   |   |── ...
    |   |   |   |── 00_40.obj
    │   │   │── ...
    

NeRF Synthetic dataset:

sh script/run_nerf_demo.sh

DTU data:

sh script/run_dtu_demo.sh

👍 Acknowledgement

Our code is built on Relightable 3DGS, we sincerely thank their efforts.

📘 Citation

If you find our work useful in your research, please be so kind to cite:

@article{gao2024mani,
  title={Mani-GS: Gaussian Splatting Manipulation with Triangular Mesh},
  author={Gao, Xiangjun and Li, Xiaoyu and Zhuang, Yiyu and Zhang, Qi and Hu, Wenbo and Zhang, Chaopeng and Yao, Yao and Shan, Ying and Quan, Long},
  journal={arXiv preprint arXiv:2405.17811},
  year={2024}
}