II.15 - Symmetry: a context free Cue for Foveated Vision

Abstract

This chapter introduces a low level operator that can be used to locate the regions of high symmetry value and use them as fixation cues. However, a powerful tool for object recognition can be designed using semantics in combination with the symmetry operators. For example, a face using the circular symmetry operator, CSn (p) applied to a rough resolution of the picture, can be located and then both circular symmetry, CSn (p), and symmetry along horizontal lines, Sn (p, 1), can be applied on smaller focused regions of the picture in higher resolution. Recent object recognition paradigms have shown that recognition can be performed using interest points in unsegmented scenes. Such paradigms may use as input the symmetry operator output. Moreover, the symmetry operator produces more than interest points. One may view the computation of the symmetry measure S (p, ψ) as a symmetry edge extraction procedure, where the response in given directions is computed for all the points; the consequence of this approach is in the observation that one may apply most of the standard (for example Hough Transform, edge finking) computer vision operations to this map exactly in the same way it is applied to a standard edge map, thus enabling the extraction of more complex symmetry lines and symmetry curves.