Shared Segmentation of Natural Scenes Using Dependent Pitman-Yor Processes
Erik B. Sudderth, Michael I. Jordan · 2008
We develop a statisticalframework forthe simultaneous, unsupervised segmentation and discovery of visual object categories from image databases. Examining a large set of manually segmented scenes, we show that object frequencies and segmentsizesbothfollowpowerlawdistributions,whicharewellmodeledbythe Pitman–Yor (PY) process. This nonparametric prior distribution leads to learning algorithms which discover an unknown set of objects, and segmentation methods which automatically adapt their resolution to each image. Generalizing previousapplicationsofPYprocesses,weuseGaussianprocessestodiscoverspatially contiguous segments which respect image boundaries. Using a novel family of variationalapproximations,ourapproachproducessegmentationswhichcompare favorablytostate-of-the-artmethods,whilesimultaneouslydiscoveringcategories shared among natural scenes. 1