Deformation-Guided Unsupervised Non-Rigid Shape Matching

Aymen Merrouche, Joao Regateiro, Stefanie Wuhrer, Edmond Boyer.

arXiv Preprint

Published Version (with a short video, poster and supplementary)

code

Figure: Sample Results.

Figure: Method overview. The method takes as input a pair of 3D meshes $\mathcal{X}, \mathcal{Y}$, decomposed into a hierarchy of non-overlapping surface patches. It is divided into two steps that work from coarse to fine levels, sequentially per level. An association step extracts coarse-to-fine correspondences between $\mathcal{X}$ and $\mathcal{Y}$ as inter-patch association matrices (from $\Pi^L$ to $\Pi^0$), which induce spatial deformation targets. A deformation step outputs deformations that achieve these targets in 3D (from $\mathcal{X}_L$, $\mathcal{Y}_L$ to $\mathcal{X}_0$, $\mathcal{Y}_0$).