MASt3R-SLAM

A real-time monocular dense SLAM system built on MASt3R as a prior, able to map and localize from ordinary uncalibrated video.

MASt3R-SLAM was proposed by Andrew Davison’s group at Imperial College London (Riku Murai, Eric Dexheimer, and colleagues), published at CVPR 2025 as a Highlight paper. It is a real-time monocular dense SLAM (simultaneous localization and mapping) system designed from the ground up around MASt3R, a two-view 3D reconstruction and matching model: for each new frame, it uses MASt3R to predict the point map and dense correspondences between that frame and the keyframes, then performs camera tracking, local fusion, loop closure, and a second-order global optimization, yielding a globally consistent camera trajectory and dense 3D geometry at about 15 frames per second. Traditional monocular SLAM usually needs the camera intrinsics calibrated beforehand; this system only assumes a single optical center and doesn’t depend on a fixed parametric camera model, so it can run on uncalibrated video too — and with known calibration, a small modification lets it reach the state of the art at the time. The open-source code supports live RealSense input, MP4 video, and folders of images.

ExampleRecording a video indoors with a phone, with no camera calibration done, MASt3R-SLAM can estimate the camera trajectory and reconstruct a dense point cloud of the room.

Also called

MASt3R-SLAM: Real-Time Dense SLAM with 3D Reconstruction Priors

Related

MASt3R · DUSt3R · Simultaneous Localization and Mapping · Visual SLAM · DROID-SLAM · Pointmap

Sources

arXiv: MASt3R-SLAM: Real-Time Dense SLAM with 3D Reconstruction Priors
GitHub: rmurai0610/MASt3R-SLAM

As of

2025-06