Maintenance of Figure-Ground Segmentation by Cue-Selection

Peter Nordlund

Abstract

An approach to figure-ground segmentation based on a conjunction of motion and depth is presented. The main idea is to produce a 2-dimensional histogram with depth in one dimension and horizontal motion in the other. This histogram is then analyzed. The most significant peaks in the histogram are backprojected to the image to produce an object mask. This object mask is maintained over time. A cue-selection algorithm using both the depth-motion algorithm and another featurebased algorithm is also demonstrated. This combined algorithm manages to maintain the object mask although the depth-motion algorithm alone fails. Keywords: stereo disparity, image motion, scale-space, segmentation, cue-integration

Full paper: (PostScript 5.1M)

Peter Nordlund (petern@nada.kth.se)