An experiment on the theory of visual motion perception

Two classes of theories of motion perception were studied: correlation and gradient models (in the sense of D. Marr). Random-dot kinematograms with a shifted square were presented to subjects, and the 80% threshold for detection of correct direction of movement was determined. Correlation models predict scale invariance, that is, a constant shift measured in units of texture elements of the translated pattern in spite of a geometric magnification. This was refuted for all subjects. The increase of the recognizable translation with the area of the translated form was verified except for the patterns with the largest texture elements (8 minutes of arc). This prediction, however, is not very specific for competing theories. Gradient models are not yet fully specified, and they contain some free parameters. They cannot be tested strictly, but there exist reasonable numerical parameter values by which our data can be satisfactorily explained.