Abstract
We propose a method to learn succinct hierarchical linguistic descriptions of visual datasets, which allow for improved navigation efficiency in image collections. Classic exploratory data analysis methods, such as agglomerative hierarchical clustering, only provide a means of obtaining a tree-structured partitioning of the data. This requires the user to go through the images first, in order to reveal the semantic relationship between the different nodes. On the other hand, in this work we propose to learn a hierarchy of linguistic descriptions, referred to as attributes, which allows for a textual description of the semantic content that is captured by the hierarchy. Our approach is based on a generative model, which relates the attribute descriptions associated with each node, and the node assignments of the data instances, in a probabilistic fashion. We furthermore use a nonparametric Bayesian prior, known as the tree-structured stick breaking process, which allows for the structure of the tree to be learned in an unsupervised fashion. We also propose appropriate performance measures, and demonstrate superior performance compared to other hierarchical clustering algorithms