Register Any Point: Scaling 3D Point Cloud Registration by Flow Matching

Oral · Best Paper Candidate (top 10)

Abstract

Point cloud registration aligns multiple unposed point clouds into a common frame, a core capability for 3D reconstruction and robot localization. We cast registration as conditional generation: a learned continuous point-wise velocity field transports noisy points to a registered scene, from which the pose of each view is recovered. The resulting model performs strongly across pairwise and multi-view registration benchmarks, particularly under low overlap, and generalizes across scales and sensor modalities.