Self-Supervised Underwater Caustics Removal and Descattering via Deep Monocular SLAM

1 min read

This paper removes caustics (the moving bright patterns sunlight casts on shallow seafloor) and backscatter from underwater images, without labels. Two networks do the work, one for single-image caustics removal and one for backscatter, and both are trained with self-supervision from a deep monocular SLAM system run over a collection of underwater videos. No ground-truth colour images or caustics labels are needed, and correction runs in real time.

Findings

  • Caustics removal performs about as well as supervised methods (abstract, via the ECCV listing).
  • Colour restoration and caustics removal improve structure-from-motion keypoint matching compared with a range of other methods.

Limits

  • Abstract only so far; I found no public preprint. Caustics matter in shallow sunlit water; deeper or murkier sites have other problems.

Relevance

One of the few enhancement papers that measures what it does for geometry (matching); compare summers-2025-enhancement-matching. Caustics bite in shallow sunlit water, which makes this relevant to near-shore surveys; see underwater perception.