Constructing Visual Taxonomies by Shape

We investigate the use of statistical shape measures for segmented image regions to construct taxonomies of visual similarity. It is demonstrated that without the use of a priori knowledge, cluster analysis can be used to impose structure on heterogeneous image data sets. We develop visual taxonomies to accomplish moderate classification tasks, and provide a framework for more powerful, open-ended analysis of large data sets. The power of this method is demonstrated using a visual taxonomy of textual data, which is shown to be efficient in an MDL context