Groueix Thibault ; Fisher Matthew ; Kim Vladimir G. ; Russell Bryan C. ; Aubry Mathieu

Abstract

We present a new deep learning approach for matching deformable shapes byintroducing {\it Shape Deformation Networks} which jointly encode 3D shapes andcorrespondences. This is achieved by factoring the surface representation into(i) a template, that parameterizes the surface, and (ii) a learnt globalfeature vector that parameterizes the transformation of the template into theinput surface. By predicting this feature for a new shape, we implicitlypredict correspondences between this shape and the template. We show that thesecorrespondences can be improved by an additional step which improves the shapefeature by minimizing the Chamfer distance between the input and transformedtemplate. We demonstrate that our simple approach improves on state-of-the-artresults on the difficult FAUST-inter challenge, with an average correspondenceerror of 2.88cm. We show, on the TOSCA dataset, that our method is robust tomany types of perturbations, and generalizes to non-human shapes. Thisrobustness allows it to perform well on real unclean, meshes from the the SCAPEdataset.