Learned Gridification for Efficient Processing of Point Clouds

Anonymous

Keywords: Point clouds, graph neural networks, 3DCNN

TL;DR: We provide a method for point cloud classification using Conv3D by learning a point cloud representation that lives on a low-resolution dense 3D grid.

Abstract: We provide a method for point cloud classification using Conv3D by learning a point cloud representation that lives on a low-resolution dense 3D grid. The gridification step consists of a bipartite graph convolution, which connects the input point cloud to a target grid in 3D. We show that the method achieves remarkable performance given the network size and grid resolution on which the Conv3D blocks operate.