A Bayesian approach for scene interpretation with integrated hierarchical structure

Abstract

We propose a concept for scene interpretation with integrated hierarchical structure. This hierarchical structure is used to detect mereological relations between complex objects as buildings and their parts, e. g., windows. We start with segmenting regions at many scales, arranging them in a hierarchy, and classifying them by a common classifier. Then, we use the hierarchy graph of regions to construct a conditional Bayesian network, where the probabilities of class occurrences in the hierarchy are used to improve the classification results of the segmented regions in various scales. The interpreted regions can be used to derive a consistent scene representation, and they can be used as object detectors as well. We show that our framework is able to learn models for several objects, such that we can reliably detect instances of them in other images. © 2011 Springer-Verlag.