Random Walk Network for 3D Point Cloud Classification and Segmentation

Abstract

Object classification and segmentation via point cloud are essential for mobile robot navigation and operation. A lot of researches ranging from 3D voxels, mesh gird and multi-view were proposed based on point cloud. However, an accurate point cloud classification is still a challenging problem. Inspired by multi-label classification in images and convolutional neural networks (CNN), in this paper we present a novel network, named Random Walk Network (RWNet), which directly processes raw 3D point cloud data to classify points and, as a result, segment one 3D scene. State-of-the-art methods mainly focus on the features of one point while spatial relationships are also essential in point classification. To deal with this issue, we combine both appearance features and spatial information of feature points to restrain the point cloud processing. We employ PointNet first to generate initial point labels and adopt point labels with high confidence as seeds. We then construct the similarity matrix between seeds and low-confidence-labeled points according to their structural and spatial similarity and use Random Walk to obtain the final classification. We demonstrate our method in 3D classification task in various scenes and compare with some benchmark methods. Experimental results show that RWNet has a better performance than others.