A novel GCN-based point cloud classification model robust to pose variances. (January 2022)

• Record Type: Journal Article
• Authors: Wang, Huafeng; Zhang, Yaming; Liu, Wanquan; Gu, Xianfeng; Jing, Xin; Liu, Zicheng
• Is Part Of: Pattern recognition. Volume 121(2022)
• Publication Date: 2022-01
• Subjects: Point cloud -- Pose robust -- Graph convolutional network -- Classification; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2021.108251
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: Different from the point cloud representation of the Cartesian coordinate system, a novel rotation-independent auxiliary network is proposed with the aid of the spherical coordinate system. In order to cope with the challenge of feature extraction caused by the disorder of point cloud data itself, a novel graph convolution network was proposed. In view of the particularity of point cloud data, how to effectively extract its global and local features and how to deal with the training problem of point cloud unbalanced data is also considered in this study. Abstract: Point cloud data can be produced by many depth sensors, such as Light Detection and Ranging (LIDAR) and RGB-D cameras, and they are widely used in broad applications of robotic navigation and remote-sensing for the understanding of environment. Hence, new techniques for object representation and classification based on 3D point cloud are becoming increasingly in high demand. Due to the irregularity of the object shape, the point cloud-based object recognition is a very challenging task, especially the pose variances of a point cloud will impose many difficulties. In this paper, we tackle the challenge of pose variances in object classification based on point cloud by developing a novel end-to-end pose robust graph convolutional network. Technically, we first represent the point cloud using the spherical system instead of the traditional Cartesian system for simplicity of computation and representation. Then a pose auxiliary network is constructed with an aim to estimate the pose changes in terms of rotation angles. Finally, a graph convolutional network is constructed for object classification against the pose variations of point cloud. The experimental results show the new model outperforms the existing approaches (such as PointNet and PointNet++) on the classification task when conducting experiments on both the ModelNet40 and the ShapeNetCore dataset with a series of random rotations of a 3D point cloud. Specifically, we obtain 73.02% accuracy for classification task on the ModelNet40 with delaunay triangulation algorithm, which is much better than the state of the art algorithms, such as PointNet and PointCNN.