Abstract
Traditional models for computing 3D shape from shading have focused on two physical processes that structure the pattern of luminance in an image: One based on Lambert's law that is applicable to matte surfaces, and another based on the law of reflection that is applicable to shiny surfaces. The present research was designed to investigate another possible process for structuring the pattern of light in an image based on Snell's law and the Fresnel equations that is applicable for transparent surfaces. The stimuli in this study consisted of crown glass objects with back lit illumination, so that there were no direct reflections toward the point of observation. Observers judged the 3D shapes of these objects along designated scan lines by adjusting spline curves to match the apparent profile of the surface in depth. The results revealed that these judgments were quite accurate except for a systematic underestimation of the depth profile. We have identified two possible sources of information that may be responsible for this high level of performance. Transmitted light has the greatest intensity when the outer boundary of a back lit surface faces away from the source of illumination, which is the opposite of Lambert's law. Another possible source of information arises from light that is refracted from one surface region, and then reflected from another toward the point of observation. This causes highlights along the sides of ridges and valleys, which accentuates those surface features. These findings suggest that the analysis of 3D shape from shading is much more complex than has typically been assumed by previous theoretical analyses.
Meeting abstract presented at VSS 2016