Abstract
The Modelfest project measured 16 observers' detection thresholds on 43 diverse stimuli in human foveal vision. The project's goal was to provide a wide variety of threshold data with small standard errors for independent modelers to compare approaches. After compensating for individual differences (one parameter per subject) the threshold standard errors were impressively low, with a mean SE of less than 0.3 decilogs (7%). This small SE places extreme restrictions on spatial vision models.
We previously showed (VSS, 2005) that the 24 stimuli comprising the Gabor-like sector of the Modelfest data can be very well fit by a model based on a contrast sensitivity weighted set of local Gabors plus four parameters: one specifying the transition point from full summation to probability (attentional) summation, two specifying the spatial summation slope above and below the transition, and an asymmetry parameter needed for the asymmetric summation of baguette and tiger-tail Gabors. We now extend the same model to a wider class of stimuli. Instead of local Gabors as the building blocks we use local Gaussians, appropriately placed to produce the Gabors. The local Gaussian model, using the previously found spatial pooling exponent of about 2.5, does an excellent job of fitting the previous 24 stimuli plus the three smallest Gaussians, Bessel and disk for a total of 29 stimuli. The new model achieves a substantially better fit than standard filter models, presumably because of our limited range of full spatial summation. Fitting the remaining stimuli would require assumptions about mechanism bandwidth.
This research was supported by grant EY04776-19