Learning isometry-invariant representations for point cloud analysis

Abstract

3D shape analysis has drawn broad attention due to its increasing demands in various fields. Despite that impressive performance has been achieved on several databases, most researchers focus their efforts on improving the performance of shape classification, retrieval, segmentation, etc. They neglect the fact that the disturbances, such as orientation and deformation, may impact much on the perception, restricting the capacity of generalizing to real applications where the prior of orientation and pose is often unknown. In this paper, we conduct shape analysis on point clouds and propose the point projection feature, which is rotation-invariant. Specifically, a novel architecture is designed to mine features of different levels. We adopt a PointNet-based backbone to extract global feature for the point cloud, and the graph aggregation operation to perceive local pose variance in the Euclidean space or geodesic space. An efficient key point descriptor is designed to assign each point with different response and help recognize the overall geometry. Furthermore, a novel dataset, PKUnon-rigid, is built that is composed of non-rigid 3D objects, based on which we evaluate the capacity of several mainstream methods in terms of processing non-rigid shapes. Mathematical analyses and experimental results demonstrate that the proposed method can extract isometry-invariant representations for 3D shape analysis tasks without rotation augmentation, and outperforms other state-of-the-art methods. The proposed dataset is publicly available at https://github.com/tasx0823/PKUnon-rigid.