Nonlinear PCA/ICA for the Structure from Motion Problem

Recovering both camera motion and object shape from multiple images, called structure from motion problem, is an important and essential problem in computer vision. Generally, the result of the structure from motion problem has an ambiguity represented by a three-dimensional rotation matrix. We present two kinds of specific criteria such as independence of parameters to fix the ambiguity by choosing an appropriate rotation matrix in the sense of computer vision. Once some criterion is defined, the fixing of the ambiguity is reduced to a nonlinear extension of the PCA/ICA. We examine the efficiency through synthetic experiments.