Invariance for Various Edge Types in Simple Images

This paper describes a study done to determine the color space characteristics of geometry, material, and illumination edges in simple images. To do this study, I built an image database of simple objects using the FLUFFER (Flexible Labeler used for Fundamental Edge Research) software developed at Swarthmore College [1]. After building the database, I analyzed the edges in the RGB, normalized rg, opponent, and hue color spaces according to the methods put forth by H.M.G. Stokman of the University of Amsterdam [2]. While some edges confirmed his findings, most edges contradicted them in one or more color spaces. This paper shows that geometry and illumination edges often have variance in unexpected color spaces, and that Stokman’s assumptions are often too simplistic to be applied to real world images. Introduction In simple images, the human eye can easily identify material, geometry, and illumination boundaries. Material edges arise between different objects and within objects that are composed of several materials or colors. Geometry edges are apparent at the corners of rectangular blocks and along the curvature of circular objects. Finally, illumination edges occur from shadows and highlights within objects. Many edges can be classified as combinations of these groupings where significant image changes occur. It is rare, however, to see an edge that displays a significant material, illumination, and geometry change. Because of this rarity, this paper does not include these edges in the study. H.M.G. Stokman theorizes in his paper [2] that edges can be classified into shadow or geometry, highlight, and material edges using low level attributes in different color spaces. The first color spaces that he examined were RGB, normalized rg, opponent, and hue. The theory section gives the equations for these color spaces. According to Stokman, shape and shadow edges are invariant in the normalized rg and hue colorspaces, highlight edges are invariant in the