A bottom-up and top-down optimization framework for learning a compositional hierarchy of object classes

Abstract

Summary form only given. Learning hierarchical representations of object structure in a bottom-up manner faces several difficult issues. First, we are dealing with a very large number of potential feature aggregations. Furthermore, the set of features the algorithm learns at each layer directly influences the expressiveness of the compositional layers that work on top of them. However, we cannot ensure the usefulness of a particular local feature for object class representation based solely on the local statistics. This can only be done when more global, object-wise information is taken into account. We build on the hierarchical compositional approach (Fidler and Leonardis, 2007) that learns a hierarchy of contour compositions of increasing complexity and specificity. Each composition models spatial relations between its constituent parts.