|
|
|
 |
|
Erik Sudderth
University of California, Berkeley, USA
Monday 6 October 2008
15.30
4 th Floor Seminar Room, Alexandra House
17 Queen Square, London, WC1N 3AR
Shared Segmentation of Natural Scenes using Dependent Pitman-Yor Processes
We explore statistical frameworks for the 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 segment sizes both follow power law distributions, which are poorly captured by standard methods. Motivated by this, we develop an alternative family of models based on the Pitman-Yor (PY) process, a generalization of the Dirichlet 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 previous applications of PY priors, we use non-Markov Gaussian processes to infer spatially contiguous segments which respect image boundaries. Using a novel family of variational approximations, our approach produces segmentations which compare favorably to state-of-the-art methods, while simultaneously discovering categories shared among natural scenes