GS-Blur [NeurIPS 2024 Datasets and Benchmark Track]

This is official implementation of GS-Blur: A 3D Scene-Based Dataset for Realistic Image Deblurring (NeurIPS 2024).

Download Link: [Google Drive]

WIP: Currently, we have only uploaded the miniset (=2000 images), but we will upload the full set as soon as we decide on an appropriate online storage solution.

Quickstart

Local Setup

Preprocess MVImgNet Dataset

Train 3DGS

Render Blurry-Sharp Pairs

Citation

If you find this useful, please consider citing our paper:

@inproceedings{lee2023exblurf,
    Title        = {GS-Blur: A 3D Scene-Based Dataset for Realistic Image Deblurring},
    Author       = {Lee, Dongwoo and Park, Joonkyu and Lee, Kyoung Mu},
    Booktile     = {Advances in Neural Information Processing Systems},
    Year         = {2024}
}

Acknowledge

This source code is derived from the implementation of 3D Gaussian Splatting. We appreciate the effort of the contributor to that repository.