Using Markov random fields for contour-based grouping

To overcome fragmentation of an initial contour-based segmentation and to organize contour segments into image primitives on a higher level of abstraction, regularities of the image data are exploited using ideas from the Gestalt psychology. First, groups are hypothesized within a hierarchy based on local evidence only, where the criteria are derived from a hand labelled training set. These hypotheses are subsequently judged in a global context using a Markov random field to derive a global interpretation. Examples of results for real data are given