| The goal of this dissertation is to label datapoints into two groups utilizing higher order information among them. More specifically, given likelihood (or error) measures that subsets of data are generated by a pattern (or belong to a class), we wish to label the individual datapoints into two classes. Several computer vision problems deal with data in this format. In model estimation, small subsets of data are randomly sampled to produce an error measure. Groups of object parts are often used to provide useful geometrical information for object recognition/ localization.;We propose a novel labeling algorithm by modeling the datapoints as nodes of a higher order Markov Network. A new higher order clique function plays the central role in our method. The behavior of this clique function is analyzed to explain how it affects the inference and to predict when it is theoretically guaranteed to produce an optimal solution. We also describe how a parametric form of the proposed clique function can be learned from labeled data. Results on several different computer vision problems will be presented to demonstrate the effectiveness of the proposed algorithm . |
