Abstract
A number of models of motion perception have included an expectation that image motion is likely to be slow - a slowness prior. Weiss et al. (2002, Nature Neuroscience, 5, 598–604) have shown that incorporating a slowness prior into their spatio-temporal gradient model can allow the prediction of a wide range of motion phenomena. However, for 2D patterns, such as plaids, the measured speed and direction from their technique varies with position, requiring smoothing over space. Smoothing leads to the blurring of motion boundaries. Other gradient motion models have a similar problem (Johnston et al. 1999, Proc. R. Soc Lond. B, 266, 509–518). However, by avoiding matrix inversion we can avoid the need for the slowness prior while calculating speed accurately. This leaves the effects of contrast on speed perception to be accounted for on the basis of problems of implementation (e.g. limits on temporal filter sensitivities; Johnston et al. 1999, Vision Research, 39, 3849–3854) rather than prior expectations. This Jet-based gradient method also allows an entirely new approach to 2D pattern motion computation, which nevertheless relies on exactly the same neural measurements as our earlier model. Using non-orthogonal projections of derivatives of Jets it is possible to accurately calculate the image motion of 2D patterns such as plaids uniformly at all points on the plaid. Motion illusions can then be attributed to problems of implementation. This method, like that use by Weiss et al., will not work for 1D pattern without regularisation. However, one can employ a measure of the dimensionality of the spatial pattern (again using the same Jet vectors) to switch between models if required. Interesting, this analysis of motion computation over space can also explain the variety of direction tuning found in V5/MT component and pattern cells.