Parametric measurements of optic flow by humans

Motion pervades the visual world (Marr, 1982). This is so not only because images in nature are unlikely to be static, but also because motion constitutes a rich source of information to understand those images. The strategy that the brain uses to measure motion in images has been extensively studied and many theories have been proposed. The challenge has been to design a biologically plausible theory that will predict all motion phenomena and work on real images. With this approach, Yuille and Grzywacz (1998) have proposed a theoretical framework for visual motion that accounts for most existing psychophysical and physiological experiments. The theory proposes that the visual system fits internal models to the incoming retinal data, selecting the best models and their parameters. The fit begins with a measurement stage that performs local estimates of motion, such as local velocity (Bravo & Watamaniuk, 1995). These local estimates, which are noisy (Shadlen & Newsome, 1998) and sometimes ambiguous (Movshon et al., 1985), are then clustered (in a space of measurement variables) into regions whose boundaries correspond to motion boundaries. This clustering is performed by a number of competitive processes corresponding to different motions. The theory proposes that different types of tests can compete to perform this clustering. For example, the grouping can be done by either parametric or non-parametric tests. The former test will try to detect familiar motions defined by prior statistics-of-natural-scenes models, while the latter will allow the visual system to deal with general types of motion that may never have been seen before.