WaterDescatterGS: Diverse Underwater 3D Scene Reconstruction Using Gaussian Splatting

Existing physics-based approaches for underwater 3D restoration and reconstruction commonly model backscatter as a function of scene depth and attenuation coefficients. However, they often overlook the view-dependent nature of underwater scattering across camera poses, making it difficult to separate medium degradation from intrinsic scene radiance and thereby limiting both visual and geometric accuracy. To address this, we present WaterDescatterGS, an efficient physics-integrated underwater 3DGS framework that explicitly models view-dependent water degradation at the pixel level. Leveraging the multi-view geometry inherent to Structure-from-Motion, our framework dynamically derives base per-view illumination directions from known camera rotations and a single globally optimized sun direction. To accommodate complex real-world photometric variations beyond this physical prior, a Physics-Residual Appearance Embedding is introduced to absorb unmodeled degradations via view-specific angular corrections. Ultimately, these refined per-view parameters are naturally integrated into the physical model to compute precise per-pixel scattering angles coupled with the rendered depth. During inference, a cross-view attention mechanism dynamically aggregates adjacent features to enhance robust novel view synthesis in turbid water. Extensive experiments demonstrate that, despite operating entirely without clean reference supervision, WaterDescatterGS achieves state-of-the-art performance in high-fidelity novel view synthesis, novel-view-consistent scene restoration, and geometric reconstruction accuracy.