Wat3R: Underwater 3D Geometry Learning without Annotations

2026-07-09 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

Wat3R is a novel cross-domain semi-supervised learning framework designed to adapt feed-forward 3D reconstruction models from air to challenging underwater environments. This method uniquely eliminates the need for any annotated underwater data, instead learning robust geometry representations from abundant unlabeled real underwater video footage using a teacher-student architecture. It incorporates a cross-view consistency loss, which utilizes geometric cues from other views to compensate for information degradation caused by water attenuation and scattering. Furthermore, the researchers constructed Water3D, a diverse dataset covering various water bodies and scenarios, specifically for geometric task evaluation. Experimental results demonstrate Wat3R outperforms current methods in underwater multi-view depth estimation and point cloud reconstruction.

Key takeaway

For Computer Vision Engineers developing underwater robotics or mapping systems, Wat3R offers a critical advancement. You can now achieve accurate 3D geometry reconstruction without the prohibitive cost of underwater annotations, utilizing abundant unlabeled video. Consider integrating this semi-supervised, cross-domain adaptation approach to overcome data scarcity and improve model robustness in challenging aquatic environments, potentially using the Water3D dataset for evaluation.

Key insights

Wat3R enables robust underwater 3D geometry learning without annotations by adapting models from air-to-water using unlabeled video and cross-view consistency.

Principles

• Cross-domain adaptation overcomes data scarcity.
• Unlabeled real video footage suffices for robust learning.
• Geometric cues from multiple views compensate for degradation.

Method

Wat3R employs a teacher-student architecture for cross-domain semi-supervised learning, adapting feed-forward 3D reconstruction models from air to underwater scenes, complemented by a cross-view consistency loss.

In practice

• Utilize unlabeled underwater video for training.
• Employ cross-view consistency for degraded data.
• Evaluate with the Water3D dataset.

Topics

• Underwater 3D Reconstruction
• Semi-supervised Learning
• Cross-domain Adaptation
• Depth Estimation
• Point Cloud Reconstruction
• Water3D Dataset

Code references

• LSXI7/Wat3R

Best for: AI Scientist, Research Scientist, Computer Vision Engineer

Related on AIssential

• Wat3R: Underwater 3D Geometry Learning without Annotations — Wat3R enables robust underwater 3D geometry learning using unlabeled data and cross-domain adaptation, outperforming prior methods.
• WAT3R: Feedforward Underwater 3D Reconstruction — WAT3R is a feed-forward framework that uses a neural adaptation module to improve underwater 3D reconstruction by accounting for imaging degradation.
• WAT3R: Feedforward Underwater 3D Reconstruction — WAT3R employs a feed-forward neural adaptation module to achieve robust 3D reconstruction from degraded underwater imagery.
• 3DMPE: 3D Multi-Perspective Embedding — 3DMPE offers a training-free, optimization-based approach for 3D point cloud reconstruction from partial multi-view 2D projections.
• Rein3D: Reinforced 3D Indoor Scene Generation with Panoramic Video Diffusion Models — Rein3D integrates 3D Gaussian Splatting with panoramic video diffusion to restore and refine 3D indoor scenes from sparse inputs.
• Feed-Forward 3D Scene Modeling: A Problem-Driven Perspective — Feed-forward 3D reconstruction offers efficient, generalizable scene modeling by mapping 2D inputs directly to 3D.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.