Sparse Localized Deformation Components upplementary Supporting Document

This supplementary document provides further visualizations, implementation details and analysis of convergence of our method. 1 Method Implementation Details Visualizing spatially varying sparsity Our method produces a set of sparse deformation components by analyzing an input mesh animation. Compared to methods like [Tena et al. 2011] and [Kavan et al. 2010], which in principle can also be viewed as sparse decompositions, our method finds the required sparsity and region segmentation automatically using a sparsity imposing regularization term. The other methods have a fixed sparsity per vertex. Fig. 1 visualizes the number of components that affect each vertex and thus gives a measure of sparsity, that is spatially varying on the mesh surface and indicates how complex the motions of different vertices are. The sparsity is low in highly deformable regions around the mouth (the subject is mostly speaking in the input animation) where many components are required to reconstruct the input. Choice of `1/`2 minimization algorithm Updating the sparse components C involves optimizing a sum of a smooth continuously differentiable data term plus the non-smooth convex regularizer Ω 0 50 100 150 200 iteration t 400000 600000 80000