AnchorConv: Anchor Convolution for Point Clouds Analysis

Abstract

Point cloud, especially in 3D, is a common and important data structure. Recently, many point convolution methods have been proposed for point cloud processing, in which features of each point is updated by aggregating those of its neighbor points. Though existing methods can achieve satisfactory performance for point cloud analysis on several tasks, the performance may degrade when the distribution is non-uniform. In this paper, we present a new point convolution method namely Anchor Convolution (AnchorConv) for analysis of irregular and unordered point clouds. Inspired by standard grid convolution for images, each point in a point cloud is updated by fusing information from uniformly distributed anchors instead of non-uniformly distributed neighbor points. Features of each anchor are estimated by aggregating features of its neighborhood and distance relative to the anchor. Since the estimation biases and variances of different anchors are not the same, anchors are reweighted to obtain better feature representation of the center point. Experiments on ModelNet40, ShapeNetPart, S3DIS, and Semantic3D datasets show that the proposed AnchorConv outperforms state-of-the-art methods for classification and segmentation of 3D point clouds.