SGER: MCMC Algorithms for Object Recognition

点击翻译按钮获取中文摘要

Abstract IIS-9979201Hollerbach, JohnUniversity of Utah$50,000 - 12 mos.MCMC Algorithms for Object RecognitionThis is a standard SGER award. This exploratory research is about finding clothed people in images with a mixture of probabilistic and geometric techniques. Current theories of object recognition cannot be used to recognize such complex deformable objects, particularly because they do not allow exact geometric models and are on cluttered backgrounds. In this study, a bottom-up approach based on learning will be used to assemble image primitives (defined by invariant filters) into regions that could be clothing, and then to assemble these regions into regions that could be body segments, which will be assembled further into limbs, and so on, until a region is sufficiently large to indicate that a person is present. This purely bottom-up approach will be complemented by Markov chain Monte Carlo (MCMC) techniques to modify existing groups. The intention is to obtain cases that have high posterior probability. In particular, collections of segments that do not on their own pass assembly tests will be randomly advanced past those tests anyway, if the group that results has high posterior probability.