Learning Compositional Categorization Models

摘要

This contribution proposes a compositional approach to visual object categorization of scenes. Compositions are learned from the Caltech 101 database1 and form intermediate abstractions of images that are semantically situated between low-level representations and the high-level categorization. Salient regions, which are described by localized feature histograms, are detected as image parts. Subsequently compositions are formed as bags of parts with a locality constraint. After performing a spatial binding of compositions by means of a shape model, coupled probabilistic kernel classifiers are applied thereupon to establish the final image categorization. In contrast to the discriminative training of the cat-egorizer, intermediate compositions are learned in a generative manner yielding relevant part agglomerations, i.e. groupings which are frequently appearing in the dataset while simultaneously supporting the discrimination between sets of categories. Consequently, compositionality simplifies the learning of a complex categorization model for complete scenes by splitting it up into simpler, sharable compositions. The architecture is evaluated on the highly challenging Caltech 101 database which exhibits large intra-category variations. Our compositional approach shows competitive retrieval rates in the range of 53.6 ±0.88% or, with a multi-scale feature set, rates of 57.8 ± 0.79%.