Reconstructing 3D scenes underwater is difficult because water absorbs and scatters light, limiting visibility. The paper introduces RUSplatting, a 3D Gaussian Splatting framework designed for sparse-view underwater scenes, and it has been accepted at BMVC 2025.

The method combines three ideas: decoupled learning of RGB channels guided by the physics of underwater attenuation for more accurate colour restoration; a frame interpolation strategy with an adaptive weighting scheme to handle sparse views and improve consistency; and a new loss function that reduces noise while preserving edges.

In experiments, RUSplatting consistently outperformed state-of-the-art methods, with PSNR gains up to 1.90 dB. The authors also released Submerged3D, a new deep-sea dataset, along with the code, aiming to support marine robotics and underwater visual analytics.