Abstract
Visual cortex is organized around orientation information. This is well studied for luminance, from which many aspects of perceptual organization emerge, but not for color. Rather, spatio-spectral color information is normally associated with differences of averages (over wavelength and position), which leads to color opponency or appearance models (e.g., retinex). We here combine an orientation-based approach for intensity with one for color, and ask: what organizational properties emerge from the interaction between orientation information in the intensity domain with those in the color do- main. In summary, relationships between disparate heuristics about shape, color, lighting, texture, and material are revealed. We model intensity and hue variations geometrically, with gradient flows for each defining a direction (orientation) and a magnitude at every point. (Such flows correspond to measurements from orientationally-selective and double-opponent cells in visual cortex.) A previously described algorithm generates hue flows that are either parallel or transverse to shading flows; we now extend those results to the magnitude domain. We define relative hue frequency H: intuitively, how many cycles hue rotates through as luminence traverses from black to white, with saturation fixed. It generalizes observations, such as slow luminance gradi- ents likely denote surface curvature or shadows, to include the co-variation of color. The implications are non-trivial. Our main result is that three natural domains emerge. In arbitrary units, (i) when hue frequency H ≈1 the well-known 'color-shading effect' arises when the flows are parallel; this implies comparable color-luminance changes. (ii) When H < 1 a 'color- lighting effect' arises; color changes more slowly than luminance; and (iii) when H >1 a 'color-texture effect' arises. New displays illustrate these domains for random surfaces, and all are confirmed with a standard depth- comparison task. Neurally plausible implementations of these computations are easily constructed.
Meeting abstract presented at VSS 2018