Morning Edition · Friday, November 13, 2026Published at 1:16 AM EST · New York

Map2SLAM preprint wraps a frozen streaming 3D model in a revisable map and cuts trajectory error by more than 40% against uncalibrated baselines

The method trains only a lightweight retrieval head, runs above 28 frames per second on a China-market RTX 5090D, and still trails a calibrated baseline on Oxford Spires.

Feed-forward 3D foundation models can estimate camera pose and geometry from a stream of video frames online. Their weakness is that each prediction is fixed once it is produced, so drift accumulates over long trajectories. The preprint "Ma…