Hierarchical Dilated Graph Convolutional Network for Point Cloud Analysis

Abstract

The graph convolution is an effective method for the feature extraction on point cloud. However, because of the gridding effect, the features of some neighboring points would be ignored and the insufficient information would affect subsequent tasks. In this paper, we propose hierarchical dilated graph convolutional networks(HDGCN), which is able to expand the receptive field and avoid the gridding effect by using hierarchical dilated convolutional layers. The hierarchical structure helps our method effectively expand the receptive field, without missing point features as much as possible. Experimental results show that outstanding performance was achieved in point cloud analysis tasks. Accuracy rates of 93.3%, 58.7%, and 85.4% were obtained on the classification on ModelNet40, semantic segmentation on S3DIS, and object-part segmentation on ShapeNet datasets, respectively.