LLGS: Unsupervised Gaussian Splatting for Image Enhancement and Reconstruction in Pure Dark Environment

TL;DR

This work tackles the challenge of enhancing low-light imagery for accurate 3D reconstruction using 3D Gaussian Splatting. It introduces LLGS, an unsupervised framework that replaces the vanilla color model with M-Color, enabling separate control of material color and illumination through a learned enhancement module. A gray-world–based color loss and a gradient-based sharpness term guide zero-knowledge priors to achieve multi-view consistency without ground-truth supervision, yielding improved PSNR/SSIM and feature matching on real datasets. The approach enables real-time rendering in dark environments and holds promise for robust robotic perception in challenging lighting conditions, including mines and underwater settings.

Abstract

3D Gaussian Splatting has shown remarkable capabilities in novel view rendering tasks and exhibits significant potential for multi-view optimization.However, the original 3D Gaussian Splatting lacks color representation for inputs in low-light environments. Simply using enhanced images as inputs would lead to issues with multi-view consistency, and current single-view enhancement systems rely on pre-trained data, lacking scene generalization. These problems limit the application of 3D Gaussian Splatting in low-light conditions in the field of robotics, including high-fidelity modeling and feature matching. To address these challenges, we propose an unsupervised multi-view stereoscopic system based on Gaussian Splatting, called Low-Light Gaussian Splatting (LLGS). This system aims to enhance images in low-light environments while reconstructing the scene. Our method introduces a decomposable Gaussian representation called M-Color, which separately characterizes color information for targeted enhancement. Furthermore, we propose an unsupervised optimization method with zero-knowledge priors, using direction-based enhancement to ensure multi-view consistency. Experiments conducted on real-world datasets demonstrate that our system outperforms state-of-the-art methods in both low-light enhancement and 3D Gaussian Splatting.

Figures (5)

• Figure 1: A comparative between LLGS and the SOTA low-light enhancement model NeRco, using identical multi-view inputs, shows that images (a) and (b) exhibit a error in view consistency post-NeRco processing, while images (c) and (d) underscore the robust view consistency maintained by LLGS, as emphasized in the View Compare image through detailed magnification.
• Figure 2: Pipeline of our proposed LLGS. 1) preprocess images and then initial 3DGS from SfM. 2) the similar rendering pipeline with the 3DGS but with our enhanced color expression M-Color. 3) The M-Color and rendering inputs datails are shown at the right part.
• Figure 3: Visual comparison between the SOTA low light 2D image enhancement with vanilla 3DGS and ours. For the Input, the bottom left is the actual input for the two systems, the top right corner is the image in HDR help understand the input information. We compared the results between Retinexformer, RQ-LLIE, SCI, LLNeRF [17], [15], [13], [18]. The results shows that our method demonstrates the best visual effects and the clearest image information.
• Figure 4: With the input of dark images, our method achieved the highest feature point matching accuracy.
• Figure 5: For the Input, the bottom left corner is the actual input for the two systems, the top right corner is the image in HDR. The first line is the ablation study for preprocess. The blue points represent Gaussians. The second line is the ablation study for gradient loss.