SeaSplat constrains 3D Gaussian splatting with a physically grounded underwater image formation model, so range- and colour-dependent effects of the water are explained by the medium instead of being baked into the Gaussians’ colours. Once trained, it can render new views with the water present, and restored views with it removed.
Findings
- Tested on the SeaThru-NeRF scenes, a scene collected by an underwater vehicle in the US Virgin Islands, and real scenes with simulated degradation. The abstract reports better rendering of held-out views, recovery of the true scene colour, and better depth maps, while keeping splatting’s speed.
Limits
- Abstract only so far. Assumes posed images from structure from motion.
Relevance
The claim that the image model helps learn scene structure (better depth) still needs checking. See underwater perception.
Links
- arXiv · project page
- Subfield: Field robotics, Mapping and localisation