Nonlinear transformation from threshold to suprathreshold texture similarity judgments

Suprathreshold similarity judgments can depart substantially from predictions based on thresholds. For example, judgment of suprathreshold grating contrast is more nearly veridical than predicted by the inverted-U threshold contrast sensitivity function. However, it is unclear whether the shift from threshold to suprathreshold judgments is merely one of channel-specific gain changes. To address this, we compared suprathreshold similarity judgments and thresholds in a well-characterized multidimensional domain of visual textures (Victor and Conte 2015). Suprathreshold similarity judgments were obtained with the paradigm of Waraich and Victor (2022). On each trial, subjects (N=4) ranked eight comparison stimuli in order of similarity to a central reference. Via a variant of multidimensional scaling, rankings across 1000 unique trials and 25 stimuli were used to build a geometric model of a perceptual space that accounted for the similarity judgments, which we compared to the perceptual space inferred from threshold sensitivities for the same stimuli. We found two kinds of differences. First, high-order image statistics contributed disproportionately to suprathreshold distances. That is, texture pairs that were predicted to be equally distant based on threshold sensitivities were perceived as more dissimilar if they differed in higher-order statistics (crossings and corners), vs. lower-order ones (luminance and edges). This finding implies selective gain increases for higher-order statistics. Second, coordinate axes in the threshold perceptual space became sharply bent in the suprathreshold space. Gain changes alone could not do this: with just gain changes, straight lines in the threshold space would map to straight lines in the suprathreshold space. Instead, the observed distortion implies that positive and negative values of the same image statistic, which are treated similarly in the threshold space, undergo separate transformations when used for similarity judgments. In sum, channel-specific gain changes and rectification-like nonlinearities are needed to account for the shift from threshold sensitivities to suprathreshold similarity.