Learning Top-Down Grouping of Compositional Hierarchies for Recognition

The complexity of real world image categorization and scene analysis requires compositional strategies for object representation. This contribution establishes a compositional hierarchy by first performing a perceptual bottom-up grouping of edge pixels to generate salient contour curves. A subsequent recursive top-down grouping yields a hierarchy of compositions. All entities in the compositional hierarchy are incorporated in a Bayesian network that couples them together by means of a shape model. The probabilistic model underlying top-down grouping as well as the shape model is learned automatically from a set of training images for the given categories. As a consequence, compositionality simplifies the learning of complex category models by building them from simple, frequently used compositions. The architecture is evaluated on the highly challenging Caltech 101 database1 which exhibits large intra-category variations. The proposed compositional approach shows competitive retrieval rates in the range of 53 .0 ± 0 .49%.