RecGS: Removing Water Caustic with Recurrent Gaussian Splatting

Installation

Installation generally follows vanilla Gaussian Splatting installation.

git clone [email protected]:tyz1030/recgs.git --recursive

or

git clone https://github.com/tyz1030/recgs.git --recursive

Conda environment is the same with vanilla Gaussian Splatting (original repo)

data

to do

Quickstart

conda activate gaussian_splatting
python3 train.py -s /data/xxxxxx    # train a vanila 3DGS first
python3 train_recgs.py -s /data/xxxxxx --start_checkpoint output/xxxxxx/chkpnt30000.pth
python3 render_recgs.py -s /data/xxxxxx -m output/xxxxxx

Citation

arXiv

@ARTICLE{zhang2025recgs,
  author={Zhang, Tianyi and Zhi, Weiming and Meyers, Braden and Durrant, Nelson and Huang, Kaining and Mangelson, Joshua and Barbalata, Corina and Johnson-Roberson, Matthew},
  journal={IEEE Robotics and Automation Letters}, 
  title={RecGS: Removing Water Caustic With Recurrent Gaussian Splatting}, 
  year={2025},
  volume={10},
  number={1},
  pages={668-675},
  keywords={Three-dimensional displays;Cameras;Sea floor;Filtering;Neural radiance field;Lighting;Robot vision systems;Solid modeling;Robots;Rendering (computer graphics);Deep learning for visual perception;marine robotics},
  doi={10.1109/LRA.2024.3511418}}