Predator
lastmod 2026-11-10created 2026-05-142 edits
PREDATOR is a deep learning model for registering two 3D point clouds with little overlap, presented at CVPR 2021. It exchanges features between the two point clouds through attention to predict overlap scores, then extracts correspondences from the overlap region to estimate the transformation, so it achieves a higher registration rate than existing methods even in difficult low-overlap cases. The low-overlap benchmark 3DLoMatch was proposed alongside it.
[CVPR 2021, Oral] PREDATOR: Registration of 3D Point Clouds with Low Overlap.