Deep Learning for 3D Classification Based on Point Cloud with Local Structure

Abstract

Point cloud consists of many unordered and unstructured points, which makes the simple deep learning (DL) network hard to capture the local structure of point cloud. This shortcoming limits the ability of the DL network to recognize the fine-grained features of objects. Network structure is changed in some studies for this problem, but this increases the network complexity. This paper proposes an effective preprocessing method for point cloud to deal with this problem. The local region that represents the local structure of point is searched by using a cube with fixed side length. All of the points in the local region are used to construct the feature vector of the center point located at the center of the cube. These feature vectors are input into a simple convolutional neural network. The ModelNet40 shape classification benchmark is used to evaluate the proposed method. Experimental results show that the proposed method improves the classification accuracy of the simple deep learning network.