LLSU: Joint Low-Light Enhancement and 3-D Gaussian Splatting for Underwater Visual Sensing

Abstract

Accurate visual reconstruction is crucial for autonomous underwater vehicles (AUVs), as it underpins key tasks such as dense mapping, SLAM, and object manipulation in complex deep-sea environments. Robust underwater reconstruction plays an essential role in practical marine applications such as subsea infrastructure inspection, environmental monitoring, and resource exploration, where reliable 3-D perception directly affects safety and efficiency. However, these environments pose significant challenges, with extremely low and uneven illumination severely degrading visual inputs. Traditional pipelines typically decouple image enhancement and scene reconstruction, applying single-view enhancement as a preprocessing step prior to downstream tasks. This staged approach often fails to maintain photometric consistency across views and undermines multiview geometric reasoning. To address these limitations, we propose LLSU, an end-to-end framework that jointly performs low-light enhancement and scene reconstruction, specifically tailored for low-light underwater robotic applications. Built upon 3-D Gaussian splatting (3DGS), LLSU integrates a physics-aware medium rendering module that explicitly models wavelength-dependent attenuation and backscatter, together with a gamma-controlled enhancement network that adaptively corrects nonuniform illumination in a view-consistent manner. All components are jointly optimized within a single differentiable pipeline. Extensive experiments demonstrate that LLSU improves PSNR over baseline approaches and attains 93.33% feature matching accuracy, substantially outperforming traditional modular pipelines (best baseline: 86.67%) in cross-view consistency. Our code and dataset are publicly available at https://github.com/booqo/lowlight-underwater-for-nerf-studio