UW-GS adapts Gaussian splatting to water with three additions: a colour appearance model for distance-dependent colour change, a physics-based density control step that keeps distant objects sharp, and a binary motion mask that removes moving content such as fish from the loss. Training uses a loss suited to scattering media and pseudo-depth maps.
Findings
- PSNR gains of up to 1.26 dB over existing methods (abstract).
- Releases S-UW, an underwater dataset with dynamic object masks.
Limits
- Abstract only so far. Moving things are masked out as distractors, not modelled.
Relevance
The motion mask is a small admission that underwater scenes are never static. Masking every moving pixel works for one survey, but a method that throws away everything that changed between views cannot tell passing fish from a real change in the scene. See underwater perception and its successor RUSplatting.
Links
- arXiv
- Subfield: Field robotics, Mapping and localisation