DRS-VPT: single feed-forward model hits SOTA image-to-LiDAR registration with zero-shot transfer

kwangmoo_yi · x · 2026-09-16

Fu and Fallon present DRS-VPT, a feed-forward transformer for image-to-scan registration that predicts scan pose, point maps, and coarse-to-fine feature pyramids for direct reprojective alignment. One model achieves state-of-the-art image-to-LiDAR registration in autonomous driving, competitive indoor relocalization without map-specific weights, and strong zero-shot transfer to unseen environments, unifying camera-LiDAR calibration and camera-to-map relocalization.

Related event: Oxford Researchers Unveil DRS-VPT for Direct Image-to-Point-Cloud Relocalization(2 posts)→

Original post →

More from Embodied

Embodied channel →