Maximum Likelihood Inference of 3D Structure from Image Sequences

The paper presents a new approach to recovering the 3D rigid shape of rigid objects from a 2D image sequence. The method has two distinguishing features:it exploits the rigidity of the object over the sequence of images, rather than over a pair of images; and, it estimates the 3D structure directly from the image intensity values, avoiding the common intermediate step of first estimating the motion induced on the image plane. The approach constructs the maximum likelihood (ML) estimate of all the shape and motion unknowns. We do not attempt the minimization of the ML energy function with respect to the entire set of unknown parameters. Rather, we start by computing the 3D motion parameters by using a robust factorization appraoch. Then, we refine the estimate of the object shape along the image sequence, by minimizing the ML-based energy function by a continuation-type method. Experimental results illustrate the performance of the method.