Abstract
Meshes are ubiquitous in visual computing and simulation, yet most existingmachine learning techniques represent meshes only indirectly, e.g. as the levelset of a scalar field or deformation of a template, or as a disordered trianglesoup lacking local structure. This work presents a scheme to directly generatemanifold, polygonal meshes of complex connectivity as the output of a neuralnetwork. Our key innovation is to define a continuous latent connectivity spaceat each mesh vertex, which implies the discrete mesh. In particular, our vertexembeddings generate cyclic neighbor relationships in a halfedge meshrepresentation, which gives a guarantee of edge-manifoldness and the ability torepresent general polygonal meshes. This representation is well-suited tomachine learning and stochastic optimization, without restriction onconnectivity or topology. We first explore the basic properties of thisrepresentation, then use it to fit distributions of meshes from large datasets.The resulting models generate diverse meshes with tessellation structurelearned from the dataset population, with concise details and high-quality meshelements. In applications, this approach not only yields high-quality outputsfrom generative models, but also enables directly learning challenging geometryprocessing tasks such as mesh repair.