Abstract
The water-color illusion consists of a chromatic darker wiggly contour, flanked by a lighter contour on a white background. It causes a long-range coloration effect across large areas, next to the brighter contour. The effect of coloration must be accomplished by two processes: a local process that causes spreading of the brighter contour to the region next to it, followed by a more global process that propagates the color over large areas. We examine the low-level computational processes of the first process and propose that the effect is due to data compression in the early stages of processing. We compare results, using computational compression against other operations that can give rise to smoothing --- Gaussian smoothing and anisotropic diffusion. The compression is implemented using a redundant wavelet system representation, with a smoothing, a first-order and a second-order difference kernel. Small components are disregarded and the image is reconstructed using the L0 norm. Comparative results on variations of the watercolor-illusion, with various colors, show that only compression can account for the spreading of the bright contour. Other methods will cause spreading of both dark and bright contours. Intuitively, the effect is explained as follows: Compression amounts to discarding some of the high-frequency components of small value. There are many more high-frequency components in a wiggly line than in a straight line, which explains the much stronger effect in the former. A single contour will not produce the effect, because all coefficients will have same magnitude. However, in the case of two contours of different contrast, the coefficients of the high-frequency components on the brighter edge are much smaller than the ones on the darker edge. Thus, they are reduced during compression, which in turn causes a smoothing next to the brighter edge.
Meeting abstract presented at VSS 2012