Efficient construction of joint motion-disparity neurons

Both visual motion analysis and stereo disparity perception play important roles in everyday life. Conceptually, these two important visual functions are solving similar computational problems. Visual motion is image displacement in the retina over time, while binocular disparity is image displacement over left and right retina. Numerous physiological and psychological experiments suggest visual motion and stereo perception are closely related at both the cell and behavior level. Many computational models have been proposed to account for the motion and disparity information process separably in the mammalian visual cortex. However, little effort has been devoted towards developing an integrated motion-disparity model to date. Here we develop a physiologically plausible and computationally efficient joint motion-disparity model based on the properties of receptive field profiles of cells. This model can extract the motion and disparity information simultaneous by utilizing a population of phase tuned neurons selective to both direction of motion and binocular disparity. We also prove that our integrated model can account for a family of Pulfrich effects, where an illusion of depth is created by an inter-ocular time delay. We further consider the situation when a temporal stretch rather than a pure inter-ocular time delay is introduced in the Pulfrich phenomena and provide an explanation without any modification of the model. We also describe the efficient implementation of Gabor filters on NVIDIA Graphics Processing Units (GPU) based on the triple-axis decomposition algorithm proposed by Lam and Shi, which significantly outperforms not only a PC based implementation, but also other GPU based implementation. Using this algorithm, we develop a real time joint motion-disparity system which can process the motion and disparity information at the same time. Keywords- motion, disparity, efficient, joint motion-disparity model, pulfrich effect, GP