SPEAKER: Olaya Alvarez Tunon, Postdoc at ITU.
ABSTRACT: This talk presents EIVO, a monocular visual odometry system for underwater ROV inspection, preceded by a brief introduction to factor graphs for SLAM intended for attendees without prior background in the topic.
EIVO couples foundation-model perception, namely dense optical flow and monocular depth with predicted uncertainty, with a persistent-landmark factor graph. The system rests on four design decisions, each validated by a controlled experiment over seven sequences ranging from simulation to public and proprietary ROV footage: covariance-aware point selection, persistent landmark connections, an interchangeable depth prior, and the back-end parameterisation together with its robust cost.Landmark persistence reduces the absolute trajectory error on every sequence, whereas swapping or fine-tuning the depth backbone, or replacing the robust kernel, does not affect it consistently.
The session concludes with an open discussion on the implications of these findings for the design of future visual odometry systems.