For robustness, MASt3R-SLAM performs relocalisation allowing it to handle the kidnapped robot problem.

As a purely monocular SLAM, it loses track when the camera’s view is obstructed, but as soon as the view is unblocked, it immediately relocalises and resumes mapping.

We use MASt3R's two-view prior as our only network with no fine-tuning.
By leveraging this 3D prior and making minimal assumptions on the camera model, we can handle dynamically changing zoom.

Efficient test-time optimisation and loop closure enable large-scale consistency.

Introducing MASt3R-SLAM, the first real-time monocular dense SLAM with MASt3R as a foundation.

Easy to use like DUSt3R/MASt3R, from an uncalibrated RGB video it recovers accurate, globally consistent poses & a dense map.

With @ericdexheimer.bsky.social* @ajdavison.bsky.social (*Equal Contribution)