Department of Computer Science
Speaker: Andrew Delong, University of Western Ontario
Title: Minimizing Energies with Hierarchical Costs
I'll talk about energies that contain a hierarchy of "label costs." Label costs encourage a solution (a discrete labeling) to rely on fewer unique labels. Such costs have been used in image segmentation (Zhu & Yuille), model selection (Torr), subspace detection (Lazic, Givoni, Frey & Aarabi) and facility location (Hochbaum). We recently showed how to incorporate label costs into the well-known alpha-expansion algorithm. In this talk I'll describe a class of energies with 'hierarchical costs', and show how to beat alpha-expansion's theoretical bounds with our 'hierarchical fusion' algorithm. Hierarchical costs should be useful whenever labels naturally form groups.