WaterSplatting pairs 3D Gaussian splatting for the geometry with a separate volumetric field for the water, queried once per pixel. Plain splatting renders only surfaces and so cannot represent the medium; volumetric NeRF methods can, but train slowly and do not render in real time. The split representation also lets the scene be rendered with the medium removed.
Findings
- The abstract reports better rendering quality than NeRF-based methods on the SeaThru-NeRF dataset, at real-time rendering speed.
- Later work uses it as the main baseline: Swimm3R reports +1.47 dB over it, and RUSplatting between +0.7 and +1.9 dB on three datasets.
Limits
- Abstract only so far. Needs COLMAP poses and dense, overlapping views.
Relevance
One of the five systems compared under one protocol in alvareztunon-2026-gs-underwater, which is the better place to judge it. See underwater perception and levy-2023-seathru-nerf.
Links
- arXiv · project page
- Subfield: Field robotics, Mapping and localisation