A quantitative model is developed to predict the perceived direction of moving two-dimensional patterns. The model incorporates both a simple motion energy pathway and a “texture boundary motion” pathway that incorporates response squaring before the extraction of motion energy. These pathways correspond to Fourier and non-Fourier motion pathways and are hypothesized to reflect processing in the VI-MT and V1-V2-MT pathway, respectively. A cosine-weighted sum of these pathways followed by competitive feedback inhibition accurately predicts the perceived direction for patterns composed of two cosine gratings at different orientations (“plaids”). The model also predicts direction discrimination, differences between foveal and peripheral viewing, changes in perceived direction with exposure duration, motion masking, and motion transparency.

A recent model for two-dimensional motion processing in MT has demonstrated that perceived direction can be accurately predicted by combining Fourier and non-Fourier component motion signals using a vector sum computation. The vector sum direction is computed by a neural network that weights Fourier and non-Fourier components by the cosine of the component direction relative to that of each pattern unit, after which competitive inhibition extracts the signals of the most active units. It is shown here that a minor modification of the connectivity in this network suffices to predict transitions from motion coherence to transparency under a wide range of circumstances. It is only necessary that the cosine weighting function and competitive inhibition be limited to directions within ± 120 deg of each pattern unit's preferred direction. This network responds by signaling one pattern direction for coherent motion but two distinct directions for transparent motion. Based on this, neural networks with properties of MT and MST neurons can automatically signal motion coherence or transparency. In addition, the model accurately predicts motion repulsion under transparency conditions.