Abstract

Recent psychophysical experiments have shown that the perceived direction of motion of plaids, composed of two cosine gratings moving in different directions, can deviate by up to 50° from the true direction of rigid translation. We have developed a dynamic neural network model that predicts these and other motion data. The model incorporates two motion pathways that are subsequently combined by using a vector sum operation. The first motion pathway extracts the directions of motion of the component gratings, i.e., the Fourier motion signals, while the second pathway employs filtering and full-wave rectification to extract a non-Fourier motion signal. The vector sum of these motion pathways quantitatively predicts the psychophysical data, and it explains the existence of parallel input pathways from both V1 and V2 to area MT. The model correctly predicts that non-Fourier plaids will move in the vector sum direction, and interactions across spatial scales in the model accurately predict transitions from rigid to transparent motion.

© 1992 Optical Society of America