Graph convolutional networks for 3D garment registration

Abstract

[en] The registration of a 3D mesh consists of, given two 3D objects, establishing a correspondence vector for each vertex of one mesh with one of the second one. This process can be used to obtain realistic simulations. Specifically, in this project, I will build the registration of pieces of clothing on realistic models of human bodies. To do so, I will implement an unsupervised machine learning model, using Graph Convolutional Networks, which allow working with non-euclidean data structures, such as 3D objects. This could allow us to build in a fast and efficient way the registration of the garments.