Matching the material of transparent objects: The role of background distortions

Abstract

It has been proposed that the visual system is able to estimate the refractive index of thick transparent objects from background distortions caused by them. More specifically, it was hypothesized that this is done based on a mid-level cue, the distortion field, whose computation from the input requires comparing the part of the background seen through the object with the part visible in plain view. We test two predictions derived from this hypothesis: (a) scene variables that do not change the distortion field, for instance, the density of the background texture, should not systematically influence the subjects' settings in a material matching task. (b) The uncertainty of the estimates should increase sharply, if the part of the background texture in plain view is removed. Our results are not compatible with these two predictions but are completely in line with the alternative interpretation that the subjects maximized the similarity of the distorted background textures on the image level. Additional results indicate that subjects can take relations between the distorted and the undistorted background into account if this is encouraged by the experimental design, but they do this in a simplistic way that is inappropriate to estimate the refractive index.