Lightweight Underwater Scene Reconstruction with Pose-Conditioned Neural Medium and 3D Gaussian Splatting

Abstract

Underwater 3D scene reconstruction is inherently challenging due to wavelength-dependent light absorption and scattering. SeaThru-NeRF integrates an underwater image-formation model into Neural Radiance Fields, achieving high-quality synthesis but suffering from prohibitively long training times and very low rendering throughput. 3D Gaussian Splatting (3DGS) trains in minutes and renders in real time, but lacks participating-medium modelling and produces colour distortion underwater. This work couples vanilla 3DGS with a lightweight, pose-conditioned neural medium. A compact multilayer perceptron, conditioned on camera position and view direction via sinusoidal positional encoding, predicts per-frame attenuation and backscatter coefficients that are applied as an image-space post-process, leaving the rasterisation loop unchanged. To stabilise joint optimisation, we introduce a deferred medium activation schedule and a progressive spherical-harmonics warm-up, supervised by a minimal three-term photometric loss. On the SeaThru-NeRF benchmark, the proposed method outperforms both SeaThru-NeRF and SeaSplat in PSNR, SSIM, and LPIPS averaged across four scenes, while training substantially faster than competing physics-aware approaches and rendering at real-time rates suitable for on-board AUV deployment.