Purchase this article with an account.
Hongjing Lu, Tungyou Lin, Alan Lee, Luminita Vese, Alan Yuille; Recovering the functional form of the slow-and-smooth prior in global motion perception. Journal of Vision 2010;10(7):819. doi: https://doi.org/10.1167/10.7.819.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Human motion perception has been proposed as a rational system for combining noisy measurements with prior expectations. An essential goal is to find means to experimentally discover prior distributions used by the human visual system. We aim to infer the functional form of motion prior from human performance. Stimuli consisted of 144 gratings with random orientations. Drifting velocities for signal gratings were determined by global motion, whereas those for noise gratings were randomly assigned. Observers were asked to discriminate global motion directions between a reference and a testing stimulus. In session 1, human performance was measured at ten different coherence levels, with a fixed angular difference between reference and testing direction. Session 2 measured performance but for ten angular differences with 0.7 coherence ratio. The priors included slowness, first-order and second-order smoothness. We focused on two functional forms for prior distributions: L2-norm (corresponding to Gaussian distribution) and L1-norm regularization (approximating Student's t distribution, whose shape has heavier tails than Gaussian). The weights of the three prior terms were estimated for each functional form to maximize the fit to human performance in the first experimental session. We found that the motion prior in the form of the Student's t distribution provided better agreement with human performance than did Gaussian priors. The recovered functional form of motion prior is consistent with objective statistics measured in natural environment. In addition, large weight values were found for the second-order smoothness terms, indicating the importance of high-order smoothness preference in motion perception. Further validation used the fitted model to predict observer performance in the second experimental session. The average accuracy difference between humans and model across ten experimental levels ranged within 3%∼8% for five subjects. This excellent predictive power demonstrates the fruitfulness of this approach.
This PDF is available to Subscribers Only