PointVGG: Graph Convolutional Network with Progressive Aggregating Features on Point Clouds

Abstract Shape classification and part segmentation are essential problems in computer vision. Although convolutional neural networks have achieved excellent performance on regular grid data, such as images, they have difficulty in accurately describing the shape information and geometric representation of point clouds because point clouds are irregular and disordered. Inspired by the convolution and pooling techniques used in images, we propose point convolution (Pconv) and point pooling (Ppool) on point clouds to learn high-level features from point clouds. Pconv obtains considerable local geometric information by magnifying receptive fields gradually. Ppool solves the disorder of point clouds similar to a symmetric function. However, in contrast to the symmetric function that directly aggregates local geometric information into a vector, Ppool acquires a more detailed local geometric representation by aggregating points progressively. A novel network, namely, PointVGG, with Pconv, Ppool, and graph structure for feature learning of point clouds, is presented and applied to object classification and part segmentation. Experiments show that PointVGG achieves state-of-the-art results on challenging benchmarks of 3D point clouds.