AquaVision3D: Physically-Consistent Medium Modelling for Underwater Gaussian Splatting

Underwater 3D scene reconstruction is challenging because wavelength-dependent attenuation, backscatter, and depth-dependent color shift cause photometric inconsistency across images. Existing medium-field methods condition on ray direction, assigning the same medium to every scene point along a ray and failing to capture spatial variation with depth. We present AquaVision3D, which instead conditions the medium on each Gaussian’s world-space position using a multi-resolution hash grid (instant-NGP), enabling spatially-varying attenuation, backscatter, and veiling-light parameters. We identify and fix a critical training collapse caused by an optimizer reset designed for MLPs, and we address channel imbalance and bounding-box sensitivity with practical tuning. Beyond outperforming vanilla 3D Gaussian Splatting (3DGS), AquaVision3D establishes a new spatial foundation for underwater medium modeling. While currently competitive with direction-conditioned methods on five of six scenes, the remaining gap explicitly reveals that angular scattering effects are complementary to spatial information, pointing toward a hybrid spatial-angular representation that our framework is well-positioned to adopt.

Recommended citation: Shubham Parab, Shaojun Bian, Yun Wu, Safa Tharib. (2026). "Underwater 3D Scene Reconstruction Using Neural Rendering: A Survey with Emphasis on 3D Gaussian Splatting Framework." 12th International Conference on Virtual Reality 2026. IEEE.
Download Paper | Download Slides