Abstract
Intrinsic graph convolution operators with differentiable kernel functionsplay a crucial role in analyzing 3D shape meshes. In this paper, we present afast and efficient intrinsic mesh convolution operator that does not rely onthe intricate design of kernel function. We explicitly formulate the order ofaggregating neighboring vertices, instead of learning weights between nodes,and then a fully connected layer follows to fuse local geometric structureinformation with vertex features. We provide extensive evidence showing thatmodels based on this convolution operator are easier to train, and canefficiently learn invariant shape features. Specifically, we evaluate ourmethod on three different types of tasks of dense shape correspondence, 3Dfacial expression classification, and 3D shape reconstruction, and show that itsignificantly outperforms state-of-the-art approaches while being significantlyfaster, without relying on shape descriptors. Our source code is available onGitHub.