A Study on Methods of Nonrigid Structure from Motion Based on Factorization Models
Abstract: Nonrigid structure from motion (NRSFM) is an important research topic in the field of computer vision. Most existing NRSFM methods are explored under the low-rank factorization model, and they assume that the unknown nonrigid structure matrix is low-rank. However for a given set of data, these methods can hardly select an appropriate rank automatically. In addition, it is hard for them to obtain an accurate estimation on the nonrigid structure in the cases of occlusions, outliers, noises, and complex deformations. Therefore, this project is intended to investigate NRSFM methods based on factorization methods: firstly, addressing the rank selection problem under the low-rank factorization model, the relationship between the model rank and the accuracy of nonrigid reconstruction is analyzed, and adaptive algorithms for rank selection are investigated. Then, several more effective NRSFM methods based on the low-rank factorization model are explored. Moreover, addressing the problem that the low-rank-factorization-model-based NRSFM methods cannot deal with complex deformations effectively, NRSFM methods based on the self-representation model are investigated according to the theory on data self-representation. In addition, in order to improve the algorithmic accuracy and reduce computational costs, we further investigate new structural priors for NRSFM, a weighting principle for the constraints of the used models, and the GPU-based algorithms for accelerating the explored NRSFM methods respectively. Finally, a robust and effective computational system for NRSFM is expected to be built.
Keywords: structure from motion; 3D reconstruction; Nonrigid object; factorization model; self-representation
Contact:
DONG Qiulei
E-mail: qldong@nlpr.ia.ac.cn