Modeling long-range interactions across the visual field in stereo correspondence

Abstract

AbstractWhen the eyes are converged, most objects in the visual scene will have a significant vertical disparity as measured at the retina. The pattern of vertical disparity across the retina is largely independent of object depth, depending mainly on the particular eye position adopted. Recently, Phillipson and Read (2010, European Journal of Neuroscience, doi:10.1111/j.1460-9568.2010.07454.x) showed that humans are better at achieving stereo correspondence when the vertical disparity field indicated infinite viewing distance, even when the physical viewing distance was just 30cm. They interpreted this as indicating that disparity encoding is optimized for long viewing distances, and is not updated to reflect changes in eye posture. Their results also indicated a significant effect of the visual periphery. Performance was better when the vertical disparity across the entire visual field was consistent with a given binocular eye position – even when this was not the eye position actually adopted – than when the vertical disparity beyond 20o eccentricity indicated a different eye position than that within 20o eccentricity. This is a surprising result, since (i) the task was to detect a target 8o in diameter, extending from 10o to 18o eccentricity, so information beyond 20o was completely irrelevant to the task, and (ii) many previous results indicate that the visual system detects and uses vertical disparity in local regions, even when the global vertical disparity field is not consistent with any single binocular eye position. Here, I show that this effect can be explained by a template-matching model in which the response of a population of disparity-detectors is compared with stored templates of the response expected to stimuli of known disparity.