SAGS: Structure-Aware 3D Gaussian Splatting

Evangelos Ververas^1,2*   Rolandos Alexandros Potamias^1,2*   Jifei Song²   Jiankang Deng^1,2   Stefanos Zafeiriou¹

¹Imperial College London, UK
²Huawei Noah’s Ark Lab, UK
^*Equal Contribution

ECCV 2024

This is the official implementation of SAGS: Structure-Aware Gaussian Splatting, a state-of-the-art Gaussian Splatting scene reconstruction method. SAGS implicitly encodes the geometry of the scene based on a local-global graph representation that facilitates the learning of complex deatails and enforces meaningful point displacements that preserve the scene's geometry. SAGS achieves state-of-the-art rendering performance and reduced storage requirements on benchmark novel-view synthesis datasets, while effectively mitigating floating artifacts and irregular distortions of previous methods and obtaining precise depth maps.

Installation

The project has been tested on a server configured with Ubuntu 22.04, cuda 12.4 and gcc 11.4.0. Similar configurations should also work, but have not been verified.

1. Clone this repo:

git clone https://github.com/eververas/SAGS.git --recursive
cd SAGS

2. Install dependencies

conda create -n sags python=3.9
conda activate sags

pip install torch==2.6.0 torchvision==0.21.0 torchaudio==2.6.0 --index-url https://download.pytorch.org/whl/cu124
pip install -r requirements.txt

Data

Download the scenes from the sources listed below and uncompress them into a data/ folder created in the project root path.

• The Tanks&Temples and Deep Blending datasets can be downloaded from 3D-Gaussian-Splatting project here.
• The MipNeRF360 scenes can be downloaded the paper author's website here.
• The BungeeNeRF dataset is available in Google Drive/百度网盘[提取码:4whv].

The data structure will be organised as follows:

data/
├── dataset_name
│   ├── scene1/
│   │   ├── images
│   │   │   ├── IMG_0.jpg
│   │   │   ├── IMG_1.jpg
│   │   │   ├── ...
│   │   ├── sparse/
│   │       └──0/
│   ├── scene2/
│   │   ├── images
│   │   │   ├── IMG_0.jpg
│   │   │   ├── IMG_1.jpg
│   │   │   ├── ...
│   │   ├── sparse/
│   │       └──0/
...

Training

For training a single scene, run:

bash ./single_train.sh scene exp_name

• scene: scene name. The available scene names can be seen by running bash ./single_train.sh
• exp_name: user-defined experiment name

The above script will store the log in outputs/dataset_name/scene_name/exp_name_cur_date_time.

For example, to train a model on the scene train with experiment name baseline, run the following:

bash ./single_train.sh train baseline

This will produce an output in outputs/tandt_db/tandt/train/baseline_cur_date_time.

Evaluation

The render test views and calculate metrics using the trained models, run:

bash ./single_test.sh scene exp_name

• scene: scene name. The available scene names can be seen by running bash ./single_test.sh
• exp_name: user-defined experiment name as defined in outputs/dataset_name/scene_name/exp_name

Citation

If you find our work helpful, please consider citing:

@inproceedings{ververas2024sags,
  title={SAGS: Structure-Aware 3D Gaussian splatting},
  author={Ververas, Evangelos and Potamias, Rolandos Alexandros and Song, Jifei and Deng, Jiankang and Zafeiriou, Stefanos},
  booktitle={European Conference on Computer Vision},
  pages={221--238},
  year={2024},
  organization={Springer}
}

LICENSE

Please follow the LICENSE of 3D-GS.

Acknowledgement

We thank the authors of 3D-GS and Scaffold-GS for their excellent work.