WFEM heat kernel: Discretization and applications to shape analysis and retrieval (Q1950207)

The heat diffusion kernel presents a successful tool in geometry processing, namely shape segmentation, comparison, shape descriptors, auto-diffusion maps, diffusion distances etc. There are several ways of representation the heat diffusion kernel such as a spectral representation, where spectrum of the Laplacian matrix with FEM, Voronoi-cot and cotangent weights is computed. The author uses so called wFEM heat kernel recently presented in his and Falcideno previous work. This type of discrete heat kernel is based on the weak formulation of the heat diffusion and its discretization with respect to linear finite elements. There are three basic methods for computation of the wFEM: the spectral approach, through the computation of the Laplace spectrum; the first order Taylor approximation, for the small values of the time parameter and the generalized Chebyshev method which are briefly described in proposed paper. On several examples author discussed the properties of wFEM such as well conditioned resulting linear system. He also compares the results using spectral and Chebyshev method and results obtained by the lumped FEM and wFEM methods. Then author introduces a time dependent scalar product that makes wFEM heat kernel self-adjoint and verifies that corresponding space of pieceswise lienar scalar functions defined on triangle mesh is Reproducing Kernel Hilbert space whose reproducing kernel is the linear FEM heat kernel. This wFEM heat kernel defines a canonical basis of corresponding Hilbert space which is useful for both shape characterization and function approximation. The wFEM is intrinsically scale covariant (without shape or kernel normalization) and scale invarinat through a normalization of the Laplacian eigenvalues.Corresponding diffusion distances are performed through the Chebyshev approximation. Obtained properties of wFEM are analogous to those that hold in the continuous case. Finally author experimentally verified that the wFEM descriptors are more robust against shape and scale changes.