Abstract

A hierarchical and compositional representation is an essential requirement for an image model effectively to represent many instances with huge complexity arising from varying geometric configurations and appearance types. In this chapter, we introduce a model, hierarchical tiling by And-Or Tree, which capitalizes on hierarchical region decomposition and region tiling to generate an exponential number of configurations with a small dictionary. We also describe an iterative learning algorithm automatically to learn the parts and their structures in a weakly supervised fashion. The proposed model and approach has been applied to two important image domains: natural scene categories and human figures. We present various experimental results to demonstrate advantages of our model in both cases.