Dynamic Graph CNN for Learning on Point Clouds

Point clouds provide a flexible geometric representation suitable for many computer graphics applications and are the raw output of most 3D data acquisition devices, and the success of CNNs on images suggests adapting their insights to point cloud data. Since point clouds inherently lack topological information, the authors design a network module called EdgeConv that recovers topology to enrich representation power for high-level tasks such as classification and segmentation. EdgeConv acts on graphs that are dynamically computed at each layer of the network, is differentiable, and can be plugged into existing architectures.

Unlike modules that operate in extrinsic space or treat each point independently, EdgeConv incorporates local neighborhood information, can be stacked or applied repeatedly to learn global shape properties, and in multi-layer systems its affinity in feature space captures semantic characteristics over potentially long distances in the original embedding. The authors demonstrate the model's performance on standard benchmarks including ModelNet40, ShapeNetPart, and S3DIS.