Payman Yadollahpour (Toyota Technological Institute at Chicago)
I will describe some of our recent work on learning systems for image segmentation using a two stage approach: given an image, we first obtain a /diverse/ set of top M most probable segmentations from a discrete probabilistic model, and then rank these using a discriminatively trained ranker that makes use of much more complex features than what could be tractably used in the initial model. The ranking model is learned to minimize the gap with the best segmentation in the set.
This approach allows for better exploration of the solution space than could be achieved by just inferring the most probable segmentation from the initial probabilistic model, and produces excellent segmentation results on a number of challenging datasets.
This is joint work with Gregory Shakhnarovich, Dhruv Batra, and Pavel Kisilev.