Unsupervised learning of stochastic AND-OR templates for object modeling

This paper presents a framework for unsupervised learning of a hierarchical generative image model called ANDOR Template (AOT) for visual objects. The AOT includes: (1) hierarchical composition as “AND” nodes, (2) deformation of parts as continuous “OR” nodes, and (3) multiple ways of composition as discrete “OR” nodes. These AND/OR nodes form the hierarchical visual dictionary. We show that both the structure and parameters of the AOT model can be learned in an unsupervised way from example images using an information projection principle. The learning algorithm consists two steps: i) a recursive Block-Pursuit procedure to learn the hierarchical dictionary of primitives, parts and objects, which form leaf nodes, AND nodes and structural OR nodes and ii) a Graph-Compression operation to minimize model structure for better generalizability, which produce additional OR nodes across the compositional hierarchy. We investigate the conditions under which the learning algorithm can identify, (i.e. recover) an underlying AOT that generates the data, and evaluate the performance of our learning algorithm through both artificial and real examples.