AbyssSplat: Active-Illumination-Aware Gaussian Splatting for Deep-Sea 3D Reconstruction

Abstract

A primary challenge in deep-sea 3D reconstruction is the strong coupling between active illumination and the participating water medium. Existing underwater 3D Gaussian Splatting (3DGS) methods typically model backscattering along pixel rays as a fixed constant, neglecting the distance-dependent attenuation of active light sources. In terms of optimization, the densification process often confounds view-space positional gradients with medium noise signals, leading to systematic errors in radiance measurements and redundant primitives. To address these challenges, we introduce AbyssSplat, the first 3D reconstruction framework specifically designed for complex deep-sea optical environments. AbyssSplat incorporates two core innovations: a Hybrid Light Field Illumination Attenuation Model, which explicitly models active light source scattering to achieve physical-domain decoupling between non-uniform illumination and scene-intrinsic radiance; and a Medium-Guided Densification Modulation Mechanism, which integrates medium transmittance and Gaussian attributes to construct a geometric confidence, applying adaptive gradient amplitude modulation to low-confidence regions to suppress overfitting of medium noise by Gaussian primitives. Evaluations on the Eiffel Tower deep-sea and SeaThru-NeRF shallow-water datasets demonstrate that AbyssSplat achieves consistent state-of-the-art performance across all scenes, with average PSNR improvements of 0.67 dB and 0.39 dB respectively over the best competing methods, alongside up to 49.5% LPIPS reduction—while requiring fewer Gaussian primitives and less training time.