Dynamic Graph CNN for Learning on Point Clouds

Point clouds provide a flexible geometric representation suitable forcountless applications in computer graphics; they also comprise the raw outputof most 3D data acquisition devices. While hand-designed features on pointclouds have long been proposed in graphics and vision, however, the recentoverwhelming success of convolutional neural networks (CNNs) for image analysissuggests the value of adapting insight from CNN to the point cloud world. Pointclouds inherently lack topological information so designing a model to recovertopology can enrich the representation power of point clouds. To this end, wepropose a new neural network module dubbed EdgeConv suitable for CNN-basedhigh-level tasks on point clouds including classification and segmentation.EdgeConv acts on graphs dynamically computed in each layer of the network. Itis differentiable and can be plugged into existing architectures. Compared toexisting modules operating in extrinsic space or treating each pointindependently, EdgeConv has several appealing properties: It incorporates localneighborhood information; it can be stacked applied to learn global shapeproperties; and in multi-layer systems affinity in feature space capturessemantic characteristics over potentially long distances in the originalembedding. We show the performance of our model on standard benchmarksincluding ModelNet40, ShapeNetPart, and S3DIS.