Deformable Graph Convolutional Networks Based Point Cloud Representation Learning

Abstract

Although the sparseness and irregularity of point cloud data have been successfully solved by deep neural networks.However,how to learn the local features of point clouds is still a challenging problem.Existing networks for point cloud representation learning have the problem of extracting features independently between points and points.To this end,a new spatial graph convolution is proposed.Firstly,an adaptive hole K-nearest neighbor algorithm is proposed when constructing the graph structure to maximize local topo-logical structure information.Secondly,the angle feature between each edge of the convolution kernel and the receptive field map is added to the convolution,which ensures more discriminative feature extraction.Finally,in order to make full use of local features,a novel graph pyramid pooling is proposed.This algorithm is tested on the standard public data sets ModelNet40 and ShapeNet,and the accuracy is 93.2% and 86.5% respectively.Experimental results show that the proposed algorithm is at a leading level in point cloud representation learning.