AtomGS: Atomizing Gaussian Splatting for High-Fidelity Radiance Field

Project page | Paper | Video | Viewer Pre-built for Windows

This repo contains the official implementation for the paper "AtomGS: Atomizing Gaussian Splatting for High-Fidelity Radiance Field". Our work builds upon 3DGS and GS Monitor codebases.

Installation

Environment

This codebase builds upon the 3DGS repository and maintains compatibility with it. Therefore, if you want to set up the repository smoothly, we strongly advise you to explore the video tutorial and the FAQ section from 3DGS. This may help you identify if your concern is a known issue and, ideally, lead you to a solution.

git clone --single-branch --branch main https://github.com/RongLiu-Leo/AtomGS.git
cd AtomGS
conda env create --file environment.yml
conda activate AtomGS

Viewer

The Pre-built Viewer for Windows can be found here. If you use Ubuntu or you want to check the viewer usage, please refer to GS Monitor.

Scripts

Prepare the data

Adjust the data into the format like:

<location>
|---input
    |---<image 0>
    |---<image 1>
    |---...

Then run

python convert.py -s <location> [--resize] #If not resizing, ImageMagick is not needed

Lauch the Viewer

<path to downloaded/compiled viewer>/bin/SIBR_remoteGaussian_app_rwdi.exe

Train an AtomGS Model

atomgs.mp4

python train.py -s <path to COLMAP or NeRF Synthetic dataset>

For object-centered or masked datasets, if there is no large background, we advise using a dense representation to achieve better geometry accuracy.

python train.py -s <path to COLMAP or NeRF Synthetic dataset> --scaling_lr 0 --iteration 7000

Some Tips

1. When facing OOM problem, try to reduce --warm_up_until
2. If you observe an oversmooth problem, try to reduce --lambda_normal; If you think the geometry is still noisy, try to increase --lambda_normal.

Important Command Line Arguments for train.py

--source_path / -s

Path to the source directory containing a COLMAP or Synthetic NeRF data set.

--model_path / -m

Path where the trained model should be stored (output/<random> by default).

--prune_threshold

Threshold is used to prune the Gaussians whose opacity falls below this value.

--clone_threshold

Threshold is used to clone the Gaussians whose positonal gradient exceeds this value.

--split_threshold

Threshold is used to split the Gaussians whose positonal gradient exceeds this value.

--atom_proliferation_until

Iteration where Atom Proliferation stops.

--warm_up_until

Iteration where warm-up strategy stops.

--lambda_ms_ssim

Influence of MS-SSIM Loss.

--lambda_normal

Influence of Edge-Aware Normal Loss.

--atom_init_quantile

The percentile of Atom Scale initialization.

View the Trained Model

python view.py -s <path to COLMAP or NeRF Synthetic dataset> -m <path to trained model> 

Important Command Line Arguments for view.py

--source_path / -s

Path to the source directory containing a COLMAP or Synthetic NeRF data set.

--model_path / -m

Path where the trained model should be stored (output/<random> by default).

--iteration

Specifies which of iteration to load.

Render the Trained Model

python render.py -s <path to COLMAP or NeRF Synthetic dataset> -m <path to trained model> 

Important Command Line Arguments for render.py

--source_path / -s

Path to the source directory containing a COLMAP or Synthetic NeRF data set.

--model_path / -m

Path where the trained model should be stored (output/<random> by default).

--render_mode

Specifies which map to render (rgb by default).

Export Poisson Mesh from the Trained Model

python export.py -s <path to COLMAP or NeRF Synthetic dataset> -m <path to GS model>

Important Command Line Arguments for export.py

--source_path / -s

Path to the source directory containing a COLMAP or Synthetic NeRF data set.

--model_path / -m

Path where the trained model should be stored (output/<random> by default).

--iteration

Specifies which of iteration to load (7000 by default).

--downsample

Downsample ratio for fusing RGB, depth, and normal maps to Poisson Mesh.

--depth_threshold

Threshold is used to cut off the background.

--poisson_depth

The maximum possible depth of the octree used in Poisson Mesh Extraction.

Citation

If you find our code or paper helps, please consider giving us a star or citing:

@misc{liu2024atomgs,
    title={AtomGS: Atomizing Gaussian Splatting for High-Fidelity Radiance Field}, 
    author={Rong Liu and Rui Xu and Yue Hu and Meida Chen and Andrew Feng},
    year={2024},
    eprint={2405.12369},
    archivePrefix={arXiv},
    primaryClass={cs.CV},
    url={https://rongliu-leo.github.io/AtomGS/}
}

License

This project is licensed under the Gaussian-Splatting License - see the LICENSE file for details.