3D data provides rich information compared to 2D images in machine vision applications. One important type of 3D data is point cloud due to its availability and flexibility. With the success of deep learning methods in almost every machine vision task, the focus of researches in point cloud processing has shifted from hand crafted shape descriptors to learned ones. Convolutional neural networks among all deep learning methods are very popular in image analysis fields, but they cannot be used for point cloud because of point cloud’s irregular format and unordered instinct. In this paper we adapted kernel correlation, a technique widely used in point clouds registration field , in order to develop a CNN-like method which extracts local information from point clouds. We propose a new way of measuring similarity between a kernel and the input point cloud data , our method demonstrates competitive results for point clouds classification task.
