Greedy Structure Learning of Hierarchical Compositional Models

In this work, we consider the problem of learning a hierarchical generative model of an object from a set of im-ages which show examples of the object in the presenceof variable background clutter. Existing approaches tothis problem are limited by making strong a-priori assump-tions about the object's geometric structure and require seg-mented training data for learning. In this paper, we pro-pose a novel framework for learning hierarchical compo-sitional models (HCMs) which do not suffer from the men-tioned limitations. We present a generalized formulation ofHCMs and describe a greedy structure learning frameworkthat consists of two phases: Bottom-up part learning andtop-down model composition. Our framework integratesthe foreground-background segmentation problem into thestructure learning task via a background model. As a result, we can jointly optimize for the number of layers in thehierarchy, the number of parts per layer and a foreground-background segmentation based on class labels only. Weshow that the learned HCMs are semantically meaningfuland achieve competitive results when compared to othergenerative object models at object classification on a stan-dard transfer learning dataset.