Evgeniy Bart (Computational Vision Lab, California Institute of Technology)
Organizing images is crucial for dealing efficiently with large image collections. In this talk, I will explore approaches to such an organization and its benefits. I introduce a non-parametric Bayesian model called TAX (similar to NCRP), which can organize images into a tree-shaped taxonomy
in an unsupervised manner.
The main conclusions are:
(a) images can be organized automatically, in a completely unsupervised manner;
(b) this organization is intuitively appealing, and helps represent and interpret images more efficiently.
The main benefits of the organization are easier navigation through image collections (for both computers
and humans) and reduced description length.
A natural question is whether a taxonomy is the optimal form of organization for natural images.
I will present experiments indicating that although taxonomies can organize images in a useful manner,
more elaborate structures may be even better suited for this task.