Abstract
The image formation process confounds surface property changes with lighting and geometric variations: are intensity variations material or shading induced? Normally it is assumed that both need be estimated concurrently, but as yet there is no physiological evidence for this in early vision. We seek, instead, to find an approximation to the inverse map that separates material from shading variations robustly (i.e., with high probability) and in a neurobiologically-plausible manner.
Formally, we focus on the spatial co-variation of shading and spectral information. When changes are material, consistent co-variation (across projected spatial coordinates) is expected; when changes are due to shading and shadowing, one expects inconsistent co-variation. We propose that consistency is defined in terms of shading and color flows, subject to the condition that color flows do not contain luminance information. We show that (i) these conditions are realized by double-opponent (red/green; blue/yellow), orientation-selective operators locally; (ii) that these operators signal iso-hue directions; and (iii) that global flows obtained from them predict important aspects of the perception of material vs. shading variation over a surface neighborhood.
This suggests a fundamental role for double-opponency very different from color edge detection: the orientationally-selective double-opponent cells in superficial V1 could be providing initial hypotheses about domain partitions within the inverse map. Double-opponent (iso-hue) flows generated in the above manner are technically the color analogue to shading flow information, and ground the initial estimation of shading and material properties.
Summary: shading flows aligned with (intensity-invariant) double-opponent color flows are robustly material.
Meeting abstract presented at VSS 2013