Exploiting Contextual Knowledge for Hybrid Classification of Visual Objects

Abstract

We consider the problem of classifying visual objects in a scene by exploiting the semantic context. For this task, we define hybrid classifiers (HC) that combine local classifiers with context constraints, and can be applied to collective classification problems (CCPs) in general. Context constraints are represented by weighted ASP constraints using object relations. To integrate probabilistic information provided by the classifier and the context, we embed our encoding in the formalism LP^MLN, and show that an optimal labeling can be efficiently obtained from the corresponding LP^MLN program by employing an ordinary ASP solver. Moreover, we describe a methodology for constructing an HC for a CCP, and present experimental results of applying an HC for object classification in indoor and outdoor scenes, which exhibit significant improvements in terms of accuracy compared to using only a local classifier.