Gaussian in the Dark 😈

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).

Requirements

• Linux OS
• NVIDIA GPUs. We experimented on A6000 GPUs (cuda 11.8).
• Python libraries: see environment.yml. You can use the following commands with Anaconda3 to create and activate your virtual environment:
  • git clone https://github.com/yec22/Gaussian-DK.git
  • cd Gaussian-DK
  • conda env create -f environment.yml
  • conda activate 3dgs_dk

Dataset

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).

Usage

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

Results

Comparison with 3DGS

Light-Up Effect

Acknowledgement

Code of this repo is rely on 3DGS, HDR-NeRF, and Pixel-GS. We thank the authors for their great job!