RobustSplat: Decoupling Densification and Dynamics for Transient-Free 3DGS

Cloning the Repository

git clone https://github.com/fcyycf/RobustSplat.git --recursive

Installation

Follow the instructions in the Gaussian Splatting to setup the environment. You can simply run the following command:

conda env create --file environment.yml

Datasets

We use the following datasets in our experiments：

NeRF On-the-go dataset

RobustNeRF dataset

Note: Since NeRF On-the-go does not include SfM data and RobustNeRF provides OpenCV-formatted camera parameters, we re-ran SfM for 3DGS-codebase requirments.

Our pre-processed datasets are available at https://huggingface.co/datasets/fcy99/RobustSplat-data.

Preparing Datasets

Preprocessing

Follow the usage in the SpotLessSplats to preprocess the dataset formats:

python ./prepare/prep_data.py --dataset <dataset path>

Structure from Motion

You should passing the sparse reconstruction and undistortion using COLMAP:

# Install the COLMAP if not already on your system
conda install conda-forge::colmap

# Run COLMAP
bash ./prepare/colmap.sh <dataset path> 

Downsample

We using factor 8 for downsample (factor 4 for patio and arcdetriomphe of NeRF On-the-go):

python ./prepare/downsample.py --dataset <dataset path> --factor <factor>

Note: The folder imges_{factor} and imges_{4*factor} will be genertated in dataset path.

Running

Run the following commands for training, rendering, and evaluation:

# Training
python train.py -s <dataset path> -m <model path> -r <factor>

# Rendering
python render.py -m <model path>

# Evaluation
python metrics.py -m <model path>

TODO List

• Release our prepared datasets.
• Release our checkpoints.

Citation

If you find this work useful, please consider citing:

@inproceedings{2025RobustSplat,
    author    = {Fu, Chuanyu and Zhang, Yuqi and Yao, Kunbin and Chen, Guanying and Xiong, Yuan and Huang, Chuan and Cui, Shuguang and Cao, Xiaochun},
    title     = {RobustSplat: Decoupling Densification and Dynamics for Transient-Free 3DGS},
    booktitle = {ICCV},
    year      = {2025}
}

Acknowledgements

This repo benefits from Gaussian Splatting, DINOv2, SpotLessSplats, and WildGaussians. Thanks for these excellent contributions.