Segmentation by Concurrent Grouping at Multiple Levels and Scales
Speaker: Stella Yu , Boston CollegeContact:
Date: October 12 2005
Time: 2:45PM to 3:45PM
Location: Patil/Kiva Seminar Room (32-G449)
Host: Mario Christoudias and Gerald Dalley, MIT CSAIL
Mario Christoudias and Gerald Dalley, 3-4278, 3-6095, cmch@csail, dalleyg@mitRelevant URL:
Objects that are either salient or familiar to viewers pop out from their background. These are two extremes of segmentation: one that can often be implemented by saliency detection on low-level features, and the other that often involves object recognition. Most segmentation scenarios, however, fall somewhere in-between. Does saliency precede recognition, or the other way around? Can a segmentation method take advantage of both shortcuts without either one dominating?
I will present two recent works regarding these questions. Our solution is to formulate segmentation as a concurrent grouping problem, where high-level object cues and low-level cues are considered simultaneously. I will also show that, when edges are examined across scales in a grouping setting, we can handle both texture segmentation and contour completion using one feature, one cue and one criterion. Benchmark results on a variety of real images are reported.
See other events that are part of MIT Machine Vision Colloquium 2005/2006
See other events happening in October 2005