Abstract
Perception of surface shape and reflectance is a difficult problem because it is underconstrained. The generalized bas-relief (GBR) ambiguity is potentially a powerful tool for probing how human vision solves this problem, as it mathematically describes an important special case of the problem of ambiguity. The GBR ambiguity shows how we can make drastic changes in the shape, reflectance, and illumination conditions of an object without changing its retinal image at all, and so it describes a fundamental obstacle to perception of shape and reflectance from shading. What statistical knowledge about real-world surfaces would help a visual system to overcome the GBR ambiguity? Do we need detailed knowledge about the statistics of real-world surfaces, or is general knowledge that mostly avoids pathological interpretations enough? Method. I examined the statistics of surface properties in digital representations of 1000 natural and man-made objects. I histogrammed local changes in orientation, reflectance, and illuminance, and also their pairwise joint distributions. I tested which of these statistical models of real-world surfaces selected the correct 3D interpretation of 2D images of objects, out of a wide range of GBR transformations of the objects. Results. Local reflectance changes tended to be small, and this simple prior was effective at selecting correct 3D interpretations of 2D images. Local orientation changes were also small, but this prior usually did not select correct 3D interpretations. Significantly, a simple prior stating that local changes in reflectance and illuminance are uncorrelated also worked very well in selecting correct 3D interpretations. Conclusion. Local statistical information about real-world surfaces can be useful in recovering 3D shape from 2D images, and general knowledge that mostly just avoids pathological interpretations can also work surprisingly well. I will also discuss psychophysical experiments that test whether these priors guide human perception of shape and reflectance from shading.
Meeting abstract presented at VSS 2015