A Bayesian framework of 2D image correspondence and 3D

UnrestrictedStructure-motion problem, the problem of estimating 3D shape of objects and their motions based on 2D image sequence, is a fundamental problem in vision research. While many biological visual systems have perception of 3D structure and motion from 2D visual stimuli, computer algorithms with similar capabilities and performance are still been sought after.; In part I of this thesis, conventional approaches to structure-motion problem were reviewed. Their underlying principles were outlined and the difficulties they face explained.; In part II, a Bayesian framework was proposed to address the structure-motion problem. First, a concept of probabilistic image correspondence inspired by difficulties facing conventional one-to-one image correspondence was proposed and formulated in a sequential Monte Carlo perspective. Then the same perspective was applied again to formulate structure-from-motion problem as a process of recursive estimation. Finally, a complete theoretical framework of structure-motion problem was constructed by combining probabilistic image correspondence and Bayesian structure-from-motion.; In part III, am implementation of this Bayesian framework was tested on artificial data. The results were subsequently examined and discussed