Evgeniy Bart (UCI/Caltech)

Title: Unsupervised learning of visual taxonomies

Abstract: A method for learning to recognize multiple visual categories simultaneously and without supervision is proposed. The method produces a hierarchical taxonomic description of the categories. The taxonomy is described by a tree, in two ways. First, each category is placed in correspondence with a distinct node of a tree; the leaves are associated to `tighter' categories, while the internal nodes are associated to broader categories. Furthermore, the leaves represent visual properties that are distinctive of the more specific categories, while internal nodes of the tree represent visual properties that are shared by categories in the corresponding subtree. Our method is based on the Nested Dirichlet Process (NDPs) which has recently been shown to be an appropriate and convenient model to organize collections of documents into distinct categories. We explore the properties of our algorithm with experiments on clustering and classification of collections of visual scenes and of object categories.

Joint work with Ian Porteous, Pietro Perona, and Max Welling