Combining Local Motion Signals: A Computational Study of Segmentation and Transparency

Abstract

Behaviourally significant information has often to be extracted from complex distributions of motion signals, which the visual system has to segment into meaningful components. Under special conditions two different motion directions are present simultaneously in the same region of the visual field: a situation called motion transparency. A closer look at the output of motion detector arrays will help us to understand what strategies may be employed in higher motion processing stages to segment the image and how complex distributions of motion directions can be represented. I used motion-defined gratings to investigate what kind of information is represented in the output of a simple two-dimensional motion detector model (2DMD). When motion-defined stripes are wide, the 2DMD output shows clearly separable regions which can easily be detected by spatial filters operating on such motion signals. When the stripes are too narrow to be resolved by such filters, the 2DMD output still reflects the presence of two motion directions, which nevertheless can be discriminated from pure noise. Only if the grain of the moving dots is below the receptive field size of the local motion detectors does the stimulus become indiscriminable from pure noise. These simulation results correspond to the psychophysical observations on segmentation and transparency, and relate well to a processing structure suggested by psychophysical and electrophysiological experiments (see Braddick and Qian, this volume). To understand motion processing we have to consider a variety of mechanisms that may serve to analyse the distributions of local motion signals, from simple spatiotemporal filters to high-level pattern recognition strategies.