The visibility of Eidolon distortions in things and stuff.

The visibility of alterations to the physical structure of images (distortions) depends on the image content and on viewing conditions. Here we measure human sensitivity to a class of image distortions, Eidolons, applied to image sets containing a range of content, from object images or scenes, to textures and materials. In an odd-one-out task with peripherally presented images, we replicate previous findings that distortions are harder to detect in images which contain large regions of texture or material and fewer segmentable object boundaries. Next, we reason that an image-computable model able to capture the critical aspects of encoding transformations should be able to predict the discriminability of distortion-image pairs, irrespective of image content. We therefore test a variety of image-computable models, treating them as perceptual metrics, using a simple hierarchical regression framework. Of the tested models, the texture statistics of the Portilla and Simoncelli model best predicted performance, beating simple Fourier-spectrum-based transforms and a biologically inspired LGN statistics model. There remains, however, a substantial gap between the best single image-computable metric and an oracle model that has information about the experimental parameters and image labels. This work compliments existing datasets in image distortion discriminability and image quality, and extends existing frameworks for comparatively evaluating the predictive performance of perceptual metrics.