Purchase this article with an account.
Garrett Hoff, Mark Brady; Human estimation of local contrast orientation in natural images. Journal of Vision 2009;9(8):1048. doi: 10.1167/9.8.1048.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Background and Objective:
Estimation of local contrast orientation is an important step in perceiving object shape. Orientation estimations may be computed via a bank of oriented filters, as in V1 hypercolumns, or by using steerable filters. Two problems become apparent when using filters to estimate orientation. First, the region of interest may contain some clutter, perturbing the filter's output. Second, filter kernels of various sizes may be used. Which scale is best? In this study, we show how human observers use one problem to solve the other.
Five subjects viewed 500 local image patches, each bisected by an object contour. Method of adjustment was used to estimate orientation. Patches were then filtered using steerable filters; the filters returning orientations and contrast magnitudes. Patches were also filtered using a weighted covariance filter, to determine the extent to which the target pattern of the steerable filter (e.g. a step edge) accounted for the luminance variance within the image patch.
Patches with high subject agreement (n=104) were analyzed further. Two algorithms predicted human estimation. In one, the steerable filter scale with maximum magnitude was chosen and the angle estimate made at that scale. In the second, the weighted covariance filter was used to select scale. If humans use variance-accounted-for in estimating orientation, we expect the second algorithm to be the better predictor of human estimation.
Results: A matched pair t-test showed that the predictions of the variance-accounted-for algorithm were more accurate than the steerable filter algorithm, t(102) = 1.679, p=.048, in predicting human estimations.
Discussion & Conclusion:
Results of this study show that the dual problems of orientation estimation can be solved by considering both at once. Observers estimate orientation by first choosing the scale that provides the greatest signal relative to the overall variance in the image patch, thus avoiding clutter.
This PDF is available to Subscribers Only