Abstract: This talk illustrates the problem of identifying regions in an image that are likely to correspond to a single object. This problem is discussed from two different points of view; first the image segmentation approach is presented, followed by a discussion of segmentation as a search problem.

In the context of search-based segmentation, an efficient algorithm for finding convex groups of line segments is presented; the motivation for this is that many objects of interest can be described by convex boundaries, or by a combination of convex parts. The proposed algorithm compares favorably to existing methods for convex group detection, and incorporates the Qualitative Probabilities framework, which enables it to select the models that best explain the scene from a pool of hundreds of different polygons.

This is a fairly high level talk, so feel free to ask about particular details you may find interesting.