An effective approach for nonrigid structure from motion with complex deformation

Abstract

In this article, an effective approach is proposed for nonrigid structure from motion to deal with the image sequence with a complex deformation. In the proposed method, a three-dimensional (3-D) structure is first divided into multiple local trajectory groups by utilizing the distance weights of the pairwise points. For each local trajectory group, a weak estimator is constructed by means of the kernel shape trajectory approach (KSTA) due to its good performance for complex deformation. In order to improve the estimation accuracy, a feature vector selection algorithm based adaptive rank selection strategy is designed to choose the approximately optimal rank parameter for KSTA. For each 3-D point, the final estimation is derived from the outputs of the weak estimators by solving a sparse optimization problem. The experimental results on several commonly used sequences demonstrate the effectiveness and feasibility of the proposed method.