Purchase this article with an account.
Joseph Catrambone, Stephen Sebastian, Peter Kvam, Tadamasa Sawada, Robert Steinman, Zygmunt Pizlo; A new approach to modeling Figure-Ground Organization. Journal of Vision 2010;10(7):1191. doi: https://doi.org/10.1167/10.7.1191.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Figure-Ground Organization (FGO) refers to finding regions and contours in a 2D image that correspond to 3D objects in a visual scene. Establishing FGO is difficult because there is not enough information in a 2D image to prevent ambiguity. The visual system (and any model) must use a priori simplicity constraints (aka priors) to eliminate these ambiguities. Doing this requires solving two problems, specifically: (i) discovering the nature of the priors, and (ii) developing operations to combine these priors with the image data. Most prior researchers assumed that because FGO was a 2-dimensional problem, 2D priors and 2D computations should be used. But, 2D priors can never convert an ill-posed problem into a well-posed problem because multiple 2D groupings and segmentations exist for any given 2D image. Not surprisingly, this approach has never worked. We treated FGO as a 3D problem and have made progress by looking for 3D symmetrical objects in the 2D image. Images of a room, containing several pieces of furniture, were used as stimuli. The model begins by selecting several starting points from which regions are grown. The growth of each region is based on a color, or color variance, similarity. Next, each region is approximated by an alpha-hull, a generalization of a convex hull. These alpha-hulls are compared to edges detected by a modified Canny edge detector. Edge detection is then performed inside each of the resulting regions. Edges are then grouped, and the cycles of edges are detected by closing the branches of a Minimum Spanning Tree produced by the edges. This produces the internal contours that are essential for detecting 3D skewed symmetry. Regions without internal contours are unlikely to correspond to 3D objects. The regions and contours found by our model are similar to those seen by the authors.
This PDF is available to Subscribers Only