CoherentGS

CoherentGS: Sparse Novel View Synthesis with Coherent 3D Gaussians
Avinash Paliwal, Wei Ye, Jinhui Xiong, Dmytro Kotovenko, Rakesh Ranjan, Vikas Chandra, Nima Khademi Kalantari
ECCV 2024

Prerequisites

You can setup the anaconda environment using:

conda env create --file environment.yml
conda activate coherentgs

CUDA 11.7 is strongly recommended.

Data Preparation

The data preprocessing scripts have now been added to the repository. You can use these to generate flow and depth for our optimization.

You can download the processed LLFF dataset here. We will add optimized pointclouds soon.

Training

Training on LLFF dataset with 3 views. You can choose from [2, 3, 4] views

python train.py --source_path path/nerf_llff_data/flower --eval --model_path output/flower --num_cameras 3

Rendering

Run the following script to render the video.

python renderpath.py -source_path path/nerf_llff_data/flower --eval --model_path output/flower

Acknowledgement

The repo is built on top of 3D Gaussian Splatting
The modified rasterizer to render depth and eval script is from FSGS

Citation

If you find our work useful for your project, please consider citing the following paper.

@inproceedings{paliwal2024coherentgs,
  title={Coherentgs: Sparse novel view synthesis with coherent 3d gaussians},
  author={Paliwal, Avinash and Ye, Wei and Xiong, Jinhui and Kotovenko, Dmytro and Ranjan, Rakesh and Chandra, Vikas and Kalantari, Nima Khademi},
  booktitle={European Conference on Computer Vision},
  pages={19--37},
  year={2024},
  organization={Springer}
}