Statistical object recognition

To be practical, recognition systems must deal with uncertainty. Positions of image features in scenes vary. Features sometimes fail to appear because of unfavorable illumination. In this work, methods of statistical inference are combined with empirical models of uncertainty in order to evaluate and refine hypotheses about the occurrence of a known object in a scene. Probabilistic models are used to characterize image features and their correspondences. A statistical approach is taken for the acquisition of object models from observations in images: Mean Edge Images are used to capture object features that are reasonably stable with respect to variations in illumination. The Alignment approach to recognition, that has been described by Huttenlocher and Ullman, is used. The mechanisms that are employed to generate initial hypotheses are distinct from those that are used to verify (and refine) them. In this work, posterior probability and Maximum Likelihood are the criteria for evaluating and refining hypotheses. The recognition strategy advocated in this work may be summarized as Align Refine Verify, whereby local search in pose space is utilized to refine hypotheses from the alignment stage before verification is carried out. Two formulations of model-based object recognition are described. MAP Model Matching evaluates joint hypotheses of match and pose, while Posterior Marginal Pose Estimation evaluates the pose only. Local search in pose space is carried out with the Expectation-Maximization (EM) algorithm. Recognition experiments are described where the EM algorithm is used to refine and evaluate pose hypotheses in 2D and 3D. Initial hypotheses for the 2D experiments were generated by a simple indexing method: Angle Pair Indexing. The Linear Combination of Views method of Ullman and Basri is employed as the projection model in the 3D experiments. (Copies available exclusively from MIT Libraries, Rm. 14-0551, Cambridge MA 02139-4307. Ph. 617-253-5668; Fax 617-253-1690.)