Computational Theories For Human Stereo Vision

Binocular stereopsis refers to the ability to perceive depth, which has always been a central problem in perception since the time of da Vinci. The foremost theoretical difficulty that arises when attempting to understand how the visual system computes disparity is known as the correspondence or matching problem. Decades of research upon macaque primary visual cortex has shown that in each layer of the primary visual cortex (V1) long-range horizontal connections among striate cortex cells exist which integrate information from different parts of the visual field. Inspired by long-range horizontal connections in V1 and the Jeffress model, a time-delay neural network which represents a time difference spatially to solve the sound localization problem, we propose a dynamic computational stereo matching algorithm that predicts how the visual system solves the stereo matching problem using left-eye and right-eye images. In our model, eye movements like saccades and drift, transform spatial information into time domain signals. A neural structure similar to the Jeffress model is used to decode disparity. To enhance performance, we introduce Gabor filters whose two-dimensional functions have been proven to be a good fit to the receptive field (RF) profiles of simple cells in the striate cortex. Further, we fitted our model with the Combination of Receptive Fields (CORF) model which is a computational model for the lateral geniculate nucleus (LGN) cell with center-surround receptive fields (RFs) proposed by Azzopardi, G., and Petkov, N. Both random-dot stereograms (RDS) and natural stereo images were used for testing. The results indicate that our model is a possible solution for the stereo matching problem, but more details need to be added for better performance