Subaquatic neural view synthesis with depth-guided refinement and multi-scale information fusion.

Abstract

Underwater images are often severely degraded by light attenuation and backscattering, which substantially impairs 3D reconstruction and novel-view synthesis. To address these challenges, we present a unified neural rendering framework for underwater novel-view synthesis. The framework couples a signed distance function (SDF) with a physics-based underwater image-formation (UIF) model. A depth-aware optimization scheme further improves geometric accuracy and color consistency. Specifically, we employ multi-resolution hash encoding for efficient spatial representation and replace analytical gradients with numerical gradients to mitigate locality bias and promote surface-normal continuity. Additionally, we introduce four pseudo-depth regularizers, namely transmittance dispersion, depth variance, inverse-depth correlation, and depth-aware frequency loss, to enhance structural stability. Extensive experiments on the SeaThru-NeRF and UW-GS shallow-water (S-UW) datasets demonstrate superior performance over prior methods across evaluation metrics, with especially large gains at long ranges and under strong scattering, indicating superior robustness and realism. These results underscore the value of physics-guided priors in neural rendering for challenging underwater conditions.