Learning with diffusion and anisotropic operators for shape correspondence

Abstract

Shape correspondence is a problem finding correspondences between two given 3d shapes. In this work, we introduce a new surface learning model, VDN, to solve the shape correspondence problem. Our model consists of two main features, diffusion and anisotropic operators. Diffusion spreads the data in the spatial dimension, so all points get the nearby information. In this manner, diffusion is used as the generalization of the convolution to curved surfaces. However, the convolution filters are restricted to be symmetric. Our model broadens the possible convolutions using the anisotropic operators from additionally learning vectors. In this way, experiments show that our method has better result than the based model. Furthermore, using only the geometric operators to communicate spatially, our model is discretization agnostic.