Abstract
This paper introduces a framework designed to accurately predict piecewiselinear mappings of arbitrary meshes via a neural network, enabling training andevaluating over heterogeneous collections of meshes that do not share atriangulation, as well as producing highly detail-preserving maps whoseaccuracy exceeds current state of the art. The framework is based on reducingthe neural aspect to a prediction of a matrix for a single given point,conditioned on a global shape descriptor. The field of matrices is thenprojected onto the tangent bundle of the given mesh, and used as candidatejacobians for the predicted map. The map is computed by a standard Poissonsolve, implemented as a differentiable layer with cached pre-factorization forefficient training. This construction is agnostic to the triangulation of theinput, thereby enabling applications on datasets with varying triangulations.At the same time, by operating in the intrinsic gradient domain of eachindividual mesh, it allows the framework to predict highly-accurate mappings.We validate these properties by conducting experiments over a broad range ofscenarios, from semantic ones such as morphing, registration, and deformationtransfer, to optimization-based ones, such as emulating elastic deformationsand contact correction, as well as being the first work, to our knowledge, totackle the task of learning to compute UV parameterizations of arbitrarymeshes. The results exhibit the high accuracy of the method as well as itsversatility, as it is readily applied to the above scenarios without anychanges to the framework.