Computational models of coherent and transparent plaid motion.

Abstract

The perceived motion of two added sinusoidal gratings of similar amplitude and spatial frequency but different orientations is often coherent. However, when either relative grating contrast or frequency are varied, perception may transform to a motion transparency. For plaids, both multiplicative and additive transparent percepts are reported. To explain perception, several computational models of motion transparency are proposed. The most general model considered is, however, a quadratic form with five unknowns. To stabilize the transparent model, additional constraints are introduced so that two velocities may be detected from the motion of plaid patterns. It is shown how this model may be realised by a two-layer (linear) feedforward network and how network learning paradigms may be used to explain some facets of visual perception. To describe the motion of plaid patterns there is an ambiguity because computational models of both coherent and transparent motion may be used to detect image velocity. In view of this competition between models, the issue of model selection is addressed; especially for cases where two or more models fit the image measurements without a residual error. The computational approach that is proposed affords one explanation why perception selects transparency in favour of coherence for plaid patterns by adjustments of relative grating contrast and frequency.