Sea-thru removes the water from an underwater photo using a range map of the scene. It builds on the authors’ revised image formation model, in which the coefficient that attenuates light from the object differs from the one that governs backscatter, and both depend on range. Backscatter is estimated from the darkest pixels at each known range; a spatially varying illuminant estimate then gives the range-dependent attenuation coefficient, and the image is inverted.
Findings
- The authors report outperforming methods built on the older atmospheric haze model, on more than 1,100 images from two optically different water bodies, which they release.
Limits
- Needs a range map per image (RGBD input), usually from structure from motion, so it inherits whatever goes wrong upstream.
- kuan-2024-sea-thru-distance finds that in simulated turbid water the method misses a 0.8 SSIM bar even with perfect depth, and needs roughly 0.30 m depth accuracy in clear water.
Relevance
The physical model that most underwater NeRF and splatting work now embeds (levy-2023-seathru-nerf, yang-2025-seasplat). It ties colour correction to geometry: wrong depth gives wrong colour. See underwater perception.
Links
- CVF open access
- Subfield: Field robotics, Perception