Rethinking Rotation Invariance with Point Cloud Registration

Recent investigations on rotation invariance for 3D point clouds have been devoted to devising rotation-invariant feature descriptors or learning canonical spaces where objects are semantically aligned. Examinations of frameworks seldom looked into. In this work, we review (RI) in terms cloud registration (PCR) and propose an effective framework via three sequential stages, namely shape encoding, aligned integration, deep registration. We first encode constructed with respect reference frames defined over different scales, e.g., local patches global topology, generate latent codes. Within the integration stage, Aligned Integration Transformer (AIT) produce a discriminative representation by integrating point-wise self- cross-relations established within Meanwhile, adopt rigid transformations between align codes consistency across scales. Finally, integrated is registered both maximize their similarities, such that preserved shared semantic information implicitly extracted from Experimental results classification, part segmentation, retrieval tasks prove feasibility our framework. Our project page released at: https://rotation3d.github.io/.