Application of Automatic Trajectory Basis Selection Method for Non-rigid Structure from Motion

Abstract

How to tune the optimal basis size is one key problem for trajectory-basis based Non-rigid Structure from Motion (NRSFM). For the small basis size, the trajectory is poorly represented; however, for the too large basis size, the trajectory system becomes more ill-conditioned and the reconstruction error becomes unbounded. As an alternation of a lot of repeated experiments, an automatic selection method is proposed to determine the optimal basis size, which is implemented by analyzing frequency spectrum of reconstruction error. Moreover, trajectory basis selection is another key problem for trajectory basis based NRSFM, Discrete Cosine Transform (DCT) is introduced as the trajectory basis. Experiments on the real Pickup image sequences are designed to test the proposed method on selecting the optimal trajectory basis to recover 3D structure of non-rigid objects. Compared with the empirical value from multiple trials, the experimental results show that the proposed automatic selection method can improve both reconstruction accuracy and efficiency of NRSFM.