RUSplatting: Robust 3D Gaussian Splatting for Sparse-View Underwater Scene Reconstruction

2 min read

RUSplatting adapts 3D Gaussian splatting to sparse, noisy deep-sea imagery. Following the revised underwater image formation model, a 5-layer MLP (deeper than the usual two) predicts per-Gaussian, per-channel attenuation and backscatter, so red, green and blue fade at their own rates with distance. To cope with few views, RIFE interpolates one frame between each pair of inputs; COLMAP is rerun on the larger set to get more starting points, and each interpolated frame gets a learnable uncertainty weight so bad interpolations count for less. An edge-aware smoothness loss uses Depth-Anything-V2 pseudo-depth to smooth noise while keeping depth edges.

Findings

  • PSNR 29.29 against 28.26 for UW-GS on the SeaThru-NeRF scenes, 25.60 against 24.88 for WaterSplatting on S-UW, and 25.80 against 24.58 for WaterSplatting on their Submerged3D set.
  • Interpolated frames raise the COLMAP starting points from 17,635 to 27,461 on average (55.7% more). Removing interpolation costs 1.84 dB PSNR in the ablation.
  • Submerged3D: four deep-sea shipwreck scenes of 20 RGB 720p images each, released on Zenodo.

Limits

  • Only rendering metrics (PSNR, SSIM, LPIPS on every eighth image held out). There is no geometric ground truth, so a better picture need not mean a better shape.
  • Depth supervision comes from Depth-Anything-V2 and the extra frames from RIFE, neither trained underwater. The authors name this as future work.
  • The grey-world loss assumes the true scene averages to grey, which is shaky in a scene dominated by one colour.
  • Offline, and needs COLMAP poses to start.

Relevance

The per-channel medium model is the right physics, but the whole pipeline stands on COLMAP succeeding, which is the first thing to fail in turbid water (alvareztunon-2026-gs-underwater). See underwater perception.