Skip to content

README.md

Latest commit

 

History

History
54 lines (38 loc) · 2.19 KB

README.md

File metadata and controls

54 lines (38 loc) · 2.19 KB

MVS-Splatting: Fast Multi-View Stereo Depth Fusion for 3D Gaussian Splatting Initialization

GitHub Pages IEEE Access

This is the code repository for the paper with the same name. We also publish a dataset called Ghent29.

Repository structure

This repository contains 4 sub-repositories, each with their own README.md.

  1. datasets: the instructions to download the datasets used in the paper.

  2. gaussian-splatting: fork of 3DGS adapted for the paper.

  3. MVSGaussian: fork of MVSGaussian (ECCV'24) adapter for the paper.

  4. MVSSplatting: our code (depth estimation etc.)

How to use

Assuming you have a dataset with the SfM (and undistortion) step of COLMAP already applied:

dataset/
├── images/                # undistorted
└── sparse/                # PINHOLE
    └── 0/
        ├── cameras.bin
        ├── images.bin
        └── points3D.bin

Use MVSSplatting to generate sparse/0/points3D_mvs.bin. This contains the 3D Gaussian Splats to initialize the training process (see gaussian-splatting) with, instead of the COLMAP SfM point cloud. During/after training, you can use 3DGS' realtime viewers to visualize the splats.

MVSSplatting also has an optional GUI.

BibTex

If you use our paper, code or dataset, please cite us. Our paper is accepted to IEEE Access, and is currently under Early Access in IEEE Xplore.

@article{mvssplatting,
  author={Artois, Julie and Lambert, Peter and Wallendael, Glenn Van},
  journal={IEEE Access}, 
  title={MVS-Splatting: Fast Multi-View Stereo Depth Fusion for 3D Gaussian Splatting Initialization}, 
  year={2025},
  volume={13},
  number={},
  pages={197796-197807},
  doi={10.1109/ACCESS.2025.3634276}}

Paper and dataset by IDLab Media