Abstract
We propose a novel approach for performing convolution of signals on curvedsurfaces and show its utility in a variety of geometric deep learningapplications. Key to our construction is the notion of directional functionsdefined on the surface, which extend the classic real-valued signals and whichcan be naturally convolved with with real-valued template functions. As aresult, rather than trying to fix a canonical orientation or only keeping themaximal response across all alignments of a 2D template at every point of thesurface, as done in previous works, we show how information across allrotations can be kept across different layers of the neural network. Ourconstruction, which we call multi-directional geodesic convolution, ordirectional convolution for short, allows, in particular, to propagate andrelate directional information across layers and thus different regions on theshape. We first define directional convolution in the continuous setting, proveits key properties and then show how it can be implemented in practice, forshapes represented as triangle meshes. We evaluate directional convolution in awide variety of learning scenarios ranging from classification of signals onsurfaces, to shape segmentation and shape matching, where we show a significantimprovement over several baselines.