The Building Blocks of Learning-based Monocular Visual Odometry for Underwater Environments

Olaya Alvarez-Tunon, Yury Brodskiy, Stella Graßhof, Andrzej Wasowski

Keywords: Deep Learning for Visual Odometry, Underwater Visual Odometry

Abstract:

Geometry-based visual odometry has long been the de facto standard due to its solid mathematical foundations, but its performance degrades in underwater environments affected by turbidity, scattering, and low contrast. Learning-based methods offer increased robustness by leveraging higher-level visual representations that are more robust to such degradations.

We revisit monocular VO as a composition of subproblems: pixel correspondence, depth estimation, and pose optimization. We propose a modular framework that combines learning-based and geometry-based components. Our approach employs neural networks for dense optical flow and monocular depth prediction, jointly estimating per-pixel uncertainties. To ensure these uncertainties are reliable and comparable across modules, we introduce a conformal prediction framework for uncertainty calibration under distribution shift.

The calibrated uncertainties are integrated into a geometry-based pose graph optimization, improving robustness and convergence. The resulting system enables flexible, modular VO design and performs robustly in visually degraded underwater conditions.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Paper Acceptance: No

Submission Number: 15