STATISTICAL APPROACHES TO FEATURE-BASED OBJECT RECOGNITION

This paper examines statistical approaches to model-based object recog nition. Evidence is presented indicating that, in some domains, normal (Gaussian) distributions are more accurate than uniform distributions for modeling feature fluctuations. This motivates the development of new maximum-likelihood and MAP recognition formulations which are base d on normal feature models. These formulations lead to an expression f or the posterior probability of the pose and correspondences given an image. Several avenues are explored for specifying a recognition hypot hesis. In the first approach, correspondences are included as a part o f the hypotheses. Search for solutions may be ordered as a combinatori al search in correspondence space, or as a search over pose space, whe re the same criterion can equivalently be viewed as a robust variant o f chamfer matching. In the second approach, correspondences are not vi ewed as being a part of the hypotheses. This leads to a criterion that is a smooth function of pose that is amenable to local search by cont inuous optimization methods. The criteria is also suitable for optimiz ation via the Expectation-Maximization (EM) algorithm, which alternate s between pose refinement and re-estimation of correspondence probabil ities until convergence is obtained. Recognition experiments are descr ibed using the criteria with features derived from video images and fr om synthetic range images.