Cite as:
Adelson E H, Weiss Y, 1998, "A simple Bayesian model predicts a complex set of motion phenomena" Perception 27 ECVP Abstract Supplement

A simple Bayesian model predicts a complex set of motion phenomena

E H Adelson, Y Weiss

To estimate the motion of an object, the visual system must combine multiple local measurements, each of which carries some degree of ambiguity. The ambiguity may arise from several sources, including the aperture problem, noise, etc. We present a model of motion perception whereby measurements from different regions are combined according to a Bayesian estimator. The estimator is one of the simplest reasonable estimators one can devise. The estimated motion maximises the posterior probability assuming a prior favouring slow and smooth velocities. We find that the estimator predicts a remarkable range of phenomena including: the bias toward vector average with plaids of low contrast, short duration, or narrow angle; the non-rigid appearance of rotating ellipses; and the strong influence of feature-like points. The model does not require specific mechanisms such as intersection of constraints, vector averaging, feature tracking, or second-order motion. It does not specify the machinery that should be used. It simply takes the raw pixel data as input and delivers the best motion estimate as output.

We argue that a range of complex phenomenology can be explained by assuming that the visual system is employing a simple and reasonable computational strategy.

These web-based abstracts are provided for ease of seaching and access, but certain aspects (such as as mathematics) may not appear in their optimum form. For the final published version of this abstract, please see
ECVP 1998 Abstract Supplement (complete) size: 2107 Kb