Abstract
Stereoscopic vision is achieved by matching images in the two eyes. It is well known that the match is easily established even when the two images significantly differ in their contrast. It is shown in this study that, given the choice between a global match to a monocular image of the same contrast or a global match to an image of a higher contrast, the higher-contrast match is preferred. This counter-intuitive result undermines correlation measures used in various stereo-matching algorithms, which minimize the difference between matched images. Instead, this preference can be described by a correlation measure maximizing a scalar product between matched images defined in multidimensional feature space. It is shown how such a correlation measure can be easily calculated based on cell types abundant in primary visual cortex.
I would like to thank Andrew Glennerster and Suzanne Mckee for their comments. This work was supported by the Welcome Trust (Grant 056657)