Séminaire MACS (Modélisation, Analyse et Calcul Scientifique).

Blanche Buet: "Geometric Inference in the Framework of Varifolds"

Europe/Paris
Description

We propose to model both regular and discrete surfaces using the concept of varifolds. Varifolds were introduced by Almgren in the study of minimal surfaces. Varifolds provide a suitable framework for the analysis of variational geometric problems: it is possible to associate a varifold structure with regular surfaces as well as with discrete analogs (such as triangulated surfaces or point clouds). This allows for a unified framework equipped with tools from varifold theory: a topology and distances for comparing two surfaces, even if one is regular and the other discrete, as well as a distributional notion of mean curvature. These ingredients, combined with a regularization step, enable the definition of a discrete curvature with convergence properties with respect to the varifold topology, typically valid when a sequence of discrete surfaces approaches a regular surface. However, the theoretical convergence rates obtained are difficult to evaluate numerically for a given sequence of point clouds, for example. We therefore propose to reconsider the question by assuming that a point cloud of N points is a realization of an i.i.d. sample (X_1, ... , X_N) drawn from a distribution supported on a surface or a d-rectifiable set S. This reformulates estimation problems (such as curvature estimation) into a statistical framework, where we aim to obtain explicit mean convergence rates (i.e., in expectation) as a function of the number of points N.