#neural-networks #cnn #point-cloud #physics

Convolutional Neural Networks (CNNs) are now a standard tool in classification and segmentation problems for computer vision. Built into CNNs are several assumptions that exploit the structure of images and array data. Recently there has been interest in adapting these networks to point cloud analysis for 3D understanding, which has applications in self driving cars and robotics (via LIDAR sensors), 3D graphics, and computer-aided design. However, many of the core assumptions of CNNs no longer hold true. This is because point clouds are not a lattice, but an unordered list of points. Many previous works attempt to map the functionality of CNNs to point clouds, but they often either simplify the convolutional structures or add a large amount of computational overhead. We discuss our efforts to give point cloud networks the same modeling capability as traditional CNNs while placing downward pressure on the overhead. We have two approaches: Graph-CNNs and Hard Directional Graph Networks. We apply these networks to standard point cloud vision benchmarks as well as segmentation of Large Hadron Collider data and Computational Fluid Dynamics simulation, where we outperform competing baselines.