PointAGCN: Adaptive Spectral Graph CNN for Point Cloud Feature Learning

Feature learning on unstructured 3D point clouds using deep networks is gaining attention due to its wide applications in robotics, self-driving and so on. Among existing methods, PointNet has achieved promising results by directly working with point cloud data. However, it does not take full advantages of neighboring points that contain fine-grained structural information which is helpful to better semantic learning. In the paper, we propose an adaptive spectral graph convolutional network for 3D point cloud feature processing, named PointAGCN. Our model use localized spectral graph convolution to capture local geometric features instead of designing of a powerful localized filter manually. The topology of the graph in each layer can be dynamically updated in each layer, which can bring more flexibility and generality. A novel graph pooling operation is carried out on the k-nearest neighbor graph, which aggregates features at graph nodes. Through extensive experiments on various datasets, the results show that the proposed approach achieves competitive performance on standard benchmarks.