CERAMATHS - DMATHS seminar: presentation by Sylvain Arguillère

The CERAMATHS mathematics department seminar will welcome Sylvain Arguillère (Laboratoire Painlevé, Université de Lille) at 2pm on Thursday, May 23, 2024, for the following talk:

Auto-encoder for 3D faces robust to mesh change

An auto-encoder is a type of algorithm using neural networks and enabling a non-linear generalization of Principal Component Analysis. These have been very successful for various applications (data standardization, compression, statistical studies, artificial data generation), particularly in the case of 2D images. However, their use in the case of 3D surfaces is more complicated, both for lack of data and, above all, because the space of meshed surfaces is far more complicated than that of images, unless we restrict ourselves to meshes with a fixed graph structure.

We propose to solve these problems by means of an auto-encoder model that is robust to mesh change (and, more generally, Lipschitz to Hausdorff distance on compacts), and by introducing a family of metrics using varifolds that allows training on data with variable topology.

This is a collaborative work with Thomas Besnier (CRIStAL), Mohamed Daoudi (CRIStAL), and Emery Pierson (Polytechnique).