A probabilistic framework for grouping image features
Rebecca Castaño, Seth Hutchinson · 2002
Presents a framework for determining probability distributions over the space of possible image feature groupings. Such a framework allows higher level processes to reason over many plausible perceptual groupings in an image, rather than committing to a specific image segmentation in the early stages of processing. The authors first derive an expression for the probability that a set of features should be grouped together, conditioned on the observed image data associated with those features. This probability measure formalizes the principle that features in an image should be grouped together when they participate in a common underlying geometric structure. The authors then present a representation scheme in which only those groupings with high probability are explicitly represented, while large sets of unlikely grouping hypotheses are implicitly represented. The authors present experimental results for a variety of real intensity images.