Scaffold-GS: Structured 3D Gaussians for View-Adaptive Rendering

Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, Bo Dai

[Project Page][arxiv]

News

[2024.04.05] Scaffold-GS is selected as a 🎈highlight in CVPR2024.

[2024.03.27] 🎈We release Octree-GS, supporting an explicit LOD representation, rendering faster in large-scale scene with high quality.

[2024.03.26] 🎈We release GSDF, which improves rendering and reconstruction quality simultaneously.

[2024.02.27] Accepted to CVPR 2024.

[2024.01.22] We add the appearance embedding to Scaffold-GS to handle wild scenes.

[2024.01.22] 🎈👀 The viewer for Scaffold-GS is available now.

[2023.12.10] We release the code.

TODO List

• Explore on removing the MLP module
• Improve the training configuration system

Overview

We introduce Scaffold-GS, which uses anchor points to distribute local 3D Gaussians, and predicts their attributes on-the-fly based on viewing direction and distance within the view frustum.

Our method performs superior on scenes with challenging observing views. e.g. transparency, specularity, reflection, texture-less regions and fine-scale details.

Installation

We tested on a server configured with Ubuntu 18.04, cuda 11.6 and gcc 9.4.0. Other similar configurations should also work, but we have not verified each one individually.

1. Clone this repo:

git clone https://github.com/city-super/Scaffold-GS.git --recursive
cd Scaffold-GS

2. Install dependencies

SET DISTUTILS_USE_SDK=1 # Windows only
conda env create --file environment.yml
conda activate scaffold_gs

Data

First, create a data/ folder inside the project path by

mkdir data

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/
...

Public Data

The BungeeNeRF dataset is available in Google Drive/百度网盘[提取码:4whv]. The MipNeRF360 scenes are provided by the paper author here. And we test on scenes bicycle, bonsai, counter, garden, kitchen, room, stump. The SfM data sets for Tanks&Temples and Deep Blending are hosted by 3D-Gaussian-Splatting here. Download and uncompress them into the data/ folder.

Custom Data

For custom data, you should process the image sequences with Colmap to obtain the SfM points and camera poses. Then, place the results into data/ folder.

Training

Training multiple scenes

To train multiple scenes in parallel, we provide batch training scripts:

• Tanks&Temples: train_tnt.sh
• MipNeRF360: train_mip360.sh
• BungeeNeRF: train_bungee.sh
• Deep Blending: train_db.sh
• Nerf Synthetic: base ->train_nerfsynthetic.sh; with warmup->train_nerfsynthetic_withwarmup.sh

run them with

bash train_xxx.sh

Notice 1: Make sure you have enough GPU cards and memories to run these scenes at the same time.

Notice 2: Each process occupies many cpu cores, which may slow down the training process. Set torch.set_num_threads(32) accordingly in the train.py to alleviate it.

Training a single scene

For training a single scene, modify the path and configurations in single_train.sh accordingly and run it:

bash ./single_train.sh

• scene: scene name with a format of dataset_name/scene_name/ or scene_name/;
• exp_name: user-defined experiment name;
• gpu: specify the GPU id to run the code. '-1' denotes using the most idle GPU.
• voxel_size: size for voxelizing the SfM points, smaller value denotes finer structure and higher overhead, '0' means using the median of each point's 1-NN distance as the voxel size.
• update_init_factor: initial resolution for growing new anchors. A larger one will start placing new anchor in a coarser resolution.

For these public datasets, the configurations of 'voxel_size' and 'update_init_factor' can refer to the above batch training script.

This script will store the log (with running-time code) into outputs/dataset_name/scene_name/exp_name/cur_time automatically.

Evaluation

We've integrated the rendering and metrics calculation process into the training code. So, when completing training, the rendering results, fps and quality metrics will be printed automatically. And the rendering results will be save in the log dir. Mind that the fps is roughly estimated by

torch.cuda.synchronize();t_start=time.time()
rendering...
torch.cuda.synchronize();t_end=time.time()

which may differ somewhat from the original 3D-GS, but it does not affect the analysis.

Meanwhile, we keep the manual rendering function with a similar usage of the counterpart in 3D-GS, one can run it by

python render.py -m <path to trained model> # Generate renderings
python metrics.py -m <path to trained model> # Compute error metrics on renderings

Viewer

The viewer for Scaffold-GS is available now.

Contact

• Tao Lu: [email protected]
• Mulin Yu: [email protected]

Citation

If you find our work helpful, please consider citing:

@article{scaffoldgs,
  author    = {Lu, Tao, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai.},
  title     = {Scaffold-GS: Structured 3D Gaussians for View-Adaptive Rendering},
  journal   = {Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2024},
}

LICENSE

Please follow the LICENSE of 3D-GS.

Acknowledgement

We thank all authors from 3D-GS for presenting such an excellent work.