Abstract
This study explores the ability of human observers to find illusory objects in various kinds of noise. The propensity to see such illusory objects in nature (often called "pareidolia") has been widely reported but has been the subject of only a handful of studies. However, recent work with deep learning networks (e.g., Deep Dreams) has shown that such networks will hallucinate such objects if one relaxes their recognition criteria. This study investigates the sensitivity and priors that humans have when searching for illusory objects in various types of noise. Method: In the first study, 24 observers were presented with 512x512 pixel noise images in which the amplitude spectrum slope varied between 1.0 and -4.0 (f1 to f-4). Observers were asked to search for a particular object (e.g. a face or a horse or a car) although such objects were never imbedded in the images. We measured the time required to find such targets in the noise (with a 30 second limit) and measured their confidence (a measure of the quality of the image they found). In the second study, we allowed observers to report any targets they found without cuing for any particular object. For all cued objects, we found that the probability of finding a target was highest and the time required to find a target was lowest when the spectra of the noise was between f-1 and f-2 (near the spectra of natural scenes). The confidence was also highest at these slopes. Faces were the fastest objects found and in general, animate objects were found faster than inanimate objects. We interpret these results in terms of the priors that observers have when finding objects in the world.
Meeting abstract presented at VSS 2016