Turning an ordinary single-camera video into a 3D model of the scene, while also figuring out where the camera was moment to moment, is a demanding job to do quickly. MAGiSt3R tackles it with a divide-and-conquer setup that runs at nearly 10 frames per second.

The idea is to have multiple agents each build local 3D point maps from chunks of the video using a fast feed-forward model, then hand those pieces to a merging component that stitches them together, both within each agent's work and across agents, into one global map. Because errors in camera position tend to pile up as you go, it adds a pose-graph optimization step to pull that accumulated drift back into line. Tested on synthetic and real footage, it reports better reconstruction and camera tracking than existing state-of-the-art methods.

This reflects only the abstract, which is light on specifics, so the paper is where the models and numbers live.