Rotation-Invariant Analysis of 3D Point Set Using Self-Attention

Abstract

Three-dimensional (3D) point set, or point cloud, is a 3D shape representation that can be captured by a 3D sensor or derived from a CAD model. Accurate analysis of 3D point sets is required for effective reuse of 3D models, or navigation and control of autonomous vehicles and robots. These 3D point sets are in general not aligned rotationally. But many applications of these 3D point sets require robustness against rotation of 3D shapes. Recent 3D point set analysis relies on Deep Neural Networks (DNN), yet most of these DNN are not robust against rotation. In this paper, we propose and evaluate a rotation invariant 3D shape analysis DNN. The DNN combines rotation normalization of local geometry using local reference frame with content adaptive feature extraction via self-attention mechanism. We evaluate the DNN on both shape classification and segmentation of 3D point sets. The proposed method is invariant to rotation about 3 axes while outperforming existing methods in accuracy.