Spectral pose transfer

Naively transferring low-frequency coefficients fails to correctly capture the reference pose and source details: (a) and (b) are respectively reference and source meshes; (c) result by copying coefficients with bases optimization (Kovnatsky et al., 2013); (d) our pose transfer. In spectral decomposition of a 3D mesh model, it is well known that eigenvectors with respect to small eigenvalues determine its main pose while eigenvectors associated with large eigenvalues encode its surface details. Based on this property, given two meshes with different connectivities, one can use coupled quasi-harmonic technique to transfer the pose of one mesh onto the other by exchanging the low-frequency coefficients in their spectral representations. However, directly synthesizing the new low frequencies with old high frequencies usually exhibits two vital artifacts: one is detail shearing and shape collapsing, and the other is medium-scale pose missing. This paper reformulates the pose transfer as a deformation problem with low-frequency coefficients as handles. It finally leads to a non-linear optimization with the coefficients as data constraint and Laplacian coordinates as regularity term for preserving details. Meanwhile, a hierarchical pose transfer framework is introduced to capture the medium-frequency poses. To reduce the computation complexity and enhance the stability we further solve the problem in a subspace defined by mean-value coordinates.