R-SLPR: Region-based Small-to-Large Point-cloud Registration with Contrastive Learning
Abstract
Point-cloud registration is fundamental to three-dimensional perception in robotic systems, but conventional approaches struggle when aligning a small, partial source point cloud to a much larger reference. R-SLPR reformulates this scale-mismatched problem as a three-stage sequence of region proposal, regional matching, and iterative refinement. It combines Fibonacci Grid Segmentation, a contrastive-learning objective, and Cascade Anchor Selection and Refinement. Evaluation on ModelNet40 establishes new state-of-the-art accuracy, reducing position and rotation mean absolute error to 0.009 and 1.104, respectively.
Type
Publication
Journal of Manufacturing Systems
Yusen Wan
Ph.D. student in Mechanical Engineering
Zeyuan Chen
M.S. student in Mechanical Engineering
Qianshi Zou
M.S. Alumni (Winter 2026)
