A Kernel Correlation-Based Approach to Adaptively Acquire Local Features for Learning 3D Point Clouds

3D models are used in a variety of CAX fields, and their key is 3D data geometry and semantic perception. However, semantic learning of 3D point clouds is a challenge due to the naturally distinct and disordered data structure, particularly for local features of point clouds. In this paper, we aim to provide machines with 3D object shape awareness, enhance the recognizability of 3D models, and enable them to allow precise geometric and semantic information in 3D point clouds. Firstly, a novel structure is proposed, namely kernel correlation learning block (KCB), which adaptively learns the local geometric features and global features at different layers, thereby enhancing the perception capacity of the network. Secondly, we developed a method to adaptively acquire and learning geometric features based on kernel correlation, and combine it with global information in the proposed KCB. Thirdly, the proposed KCB can be integrated and compatible with the typical point cloud structure in an end-to-end manner. Numerous experiments demonstrate the advantages of the proposed methods on typical 3D shape analysis approaches such as object classification, object segmentation, and semantic segmentation.