DeltaConv

Learning from 3D point-cloud data has rapidly gained momentum, motivated by the success of deep learning on images and increased availability data. In this paper, we aim to construct anisotropic convolution layers that work directly surface derived a point cloud. This is challenging because lack global coordinate system for tangential directions surfaces. We introduce DeltaConv, layer combines geometric operators vector calculus enable construction filters clouds. Because these are defined scalar- vector-fields, separate network into vector-stream, which connected operators. The stream enables explicitly represent, evaluate, process directional information. Our convolutions robust simple implement match or improve state-of-the-art approaches several benchmarks, while also speeding up training inference.