This study runs five Gaussian splatting systems with public code under one protocol (shared poses, initialisation, compute budget and evaluator) across public underwater datasets that span turbidity, loss of light and colour attenuation, plus an industrial survey. The aim is to see which advantages survive once the setup is held fixed.
Findings
- What the methods can do depends more on the setup than on the architecture (abstract).
- Water clarity binds before rendering: structure from motion registers 99.5% of frames in clear water and 0.0% at 12 NTU.
- With an artificial light that moves with the camera, medium-blind splatting beats both medium-aware systems.
- On the industrial survey, the photometric leader on the benchmark comes last, and a restoration pre-pass in front of vanilla 3DGS wins on geometry. None of this shows in the scores the field usually reports.
Limits
- Abstract only so far. Five systems, so newer methods are not all covered.
Relevance
For deciding what to trust, it says the bottleneck in murky water is getting poses at all, and that medium models built for sunlit shallow scenes can hurt under a vehicle’s own lights. See underwater perception, outlier-robust estimation, li-2025-watersplatting and yang-2025-seasplat.
Links
- arXiv · code
- Subfield: Field robotics, Robot learning, Mapping and localisation