Wat3R: Underwater 3D Geometry Learning without Annotations

2 min read

Wat3R adapts a feed-forward 3D reconstruction model from air to water without any labelled underwater data. A teacher-student setup trains on unlabelled real underwater video, and a cross-view consistency loss uses geometry from other views to make up for detail lost in the current view to attenuation and scattering. The training README says the model starts from the VGGT checkpoint. The authors also release Water3D, a benchmark covering several water bodies and scene types.

Findings

  • The abstract reports better underwater multi-view depth and point cloud reconstruction than prior methods; the README shows results as a figure, so I have no number to quote.

Limits

  • I read the abstract and README only.
  • Per the README’s data layout, each Water3D scene’s reference geometry is a COLMAP sparse model plus multi-view stereo depth maps. That reference inherits COLMAP’s failures in the murky scenes that matter most, so scores there measure agreement with COLMAP.
  • Self-training can lock in the teacher’s errors; the abstract does not say how this is guarded against.

Relevance

Wat3R needs only unlabelled dive video, a cheap route to an underwater geometry model. The evaluation problem it exposes, no trustworthy ground truth in turbid water, runs through the whole cluster; see alvareztunon-2026-gs-underwater and underwater perception.