|M.Sc Student||Peles David|
|Subject||Segmentation by Classification|
|Department||Department of Computer Science||Supervisor||Professor Michael Lindenbaum|
|Full Thesis text|
Most segmentation methods are based on a relatively simple score, designed to lend itself to relatively efficient optimization. We take the opposite approach and suggest more complex segmentation scores that are based on on-line and off-line learning processes. We train an off-line learning process to distinguish between properties of good and bad image segments. An on-line learning process, which adapts to the discriminative local features (i.e superpixels descriptors such as color, brightness, texture) is used to generate features. We embedded this score in a segmentation process which uses exploration-exploitation considerations to search a segment that maximizes the proposed score. We test our algorithm in a foreground-background segmentation task, given a minimal prior which is just a single seed point inside the object of interest. Results on GrabCut and Weizmann's image databases are presented.
Finally, a quantitative comparison of our results with the ones achieved by GrabCut and Segmentation By Composition are presented and discussed.