Dario Zanca (DIISM, University of Siena)

May 17, 2018 – 9:30 AM

DIISM, Artificial Intelligence laboratory (room 201), Siena SI

Description

Fast and cheap localization of objects of interest (OOIs) in very high resolution (VHR) images is a challenging problem. VHR can not be processed all in one step and linear search may be highly inefficient. To deal with it, I propose a scheme to learn latent policies in a hierarchy defined by the possible scales (from coarse to fine) at which input can be sampled, to process with high-resolution only certain parts of it that are good candidates for containing the OOI.

Resources

[SLIDES] Coarse-to-Fine Q-Learning
[FINAL REPORT] Coarse-to-Fine Q-Learning