This repository contains the official authors implementation associated with the paper "Gaussian in the Dark: Real-Time View Synthesis From Inconsistent Dark Images Using Gaussian Splatting", which has been accepted by Pacific Graphics 2024 (journal track).
git clone https://github.com/yec22/Gaussian-DK.git
cd Gaussian-DK
conda env create -f environment.yml
conda activate 3dgs_dk
To facilitate further research on novel view synthesis in dark conditions, we propose a new challenging dataset containing 12 real-world scenes (5 indoors and 7 outdoors). Each scene consists of approximately 80 to 130 regular format images.
The proposed dataset can be downloaded from here (~13G).
First, please make sure that all requirements are satisfied. Then, following the scripts below to train, render, and evaluate.
# Train with train/test split
python train.py -r 4 -s datasets/dark/piano -m output/piano --port 1111 --eval
# Generate renderings
python render.py -m output/piano
# Compute metrics on renderings
python metrics.py -m output/piano
# More visualization
python render_spherify.py -m output/piano
Comparison with 3DGS
Light-Up Effect
Code of this repo is rely on 3DGS, HDR-NeRF, and Pixel-GS. We thank the authors for their great job!