Using graphs for statistical object models

Signal-and knowledge-based classifiers are difficult to deploy in practical applications because of a requirement for expert knowledge elicitation and larger training data sets or robustness issues. To overcome the problems of conventional classifiers, we have been researching methods to incorporate statistical reasoning with guided object model construction for classification. Using a graph representation of object ´features,´ we model object structures statistically. The method is capable of handing different information types in a principled way. This paper covers the basic algorithm, demonstrates its application, handling occlusion and suggests future research directions.