Randomly Sparsified Synthesis for Model-Based Deformation Analysis

The tracking of deformation is one of the current challenges in computer vision. Analysis by Synthesis (AbS) based deformation tracking provides a way to fuse color and depth data into a single optimization problem very naturally. Previous work has shown that this can be done very efficiently using sparse synthesis. Although sparse synthesis allows AbS-based tracking to perform in real-time, it requires a great amount of problem specific customization and is limited to certain scenarios. This article introduces a new way of randomized adaptive sparsification of the reference model that adjusts the sparsification during the optimization process according to the required accuracy of the current optimization step. It will be shown that the efficiency of AbS can be increased significantly using the proposed method.