Abstract
What information do people use to recognize scene gist? Recent computational scene classification models have proposed that scenes' unlocalized amplitude information, their distribution of spatial frequencies and orientations, is useful for recognizing gist. Our previous research (Loschky et al., 2005; Loschky et al., 2006) has contradicted these claims, showing that phase randomization greatly hinders scene gist recognition. Experiment 1 tested the hypothesis that unlocalized amplitude information is more useful for the perceptually primitive “natural” vs. “man-made” superordinate category distinction than for the basic level distinctions previously tested, using Oliva and Torralba's (2001) 2 superordinate and 8 basic level categories. With full phase randomization, subjects were slightly above chance for basic level distinctions, but at chance for the “natural” versus “man-made” distinction. These results suggest that unlocalized amplitude information is minimally useful for making even the most primitive scene gist distinctions.
How can such results be squared with those of Guyader, et al. (2004), who showed that unlocalized amplitude information is useful for priming gist? Experiment 2 replicated their main finding, that phase randomized scenes can prime gist just as effectively as normal scenes do, in a task using two basic level categories (“coast” and “tall building”) with orthogonal dominant orientations (horizontal versus vertical), though the effect size was small. Experiment 3 tested the hypothesis that a scene's dominant orientation, while somewhat useful in priming a scene category, does not sufficiently constrain gist interpretations—there are many predominantly horizontal scenes (beaches, fields, highways, etc.), and many predominantly vertical scenes (tall buildings, trees, etc.). The results were generally consistent with this idea, with phase randomized “tree trunks” and phase randomized “tall buildings” both priming “tall building” scenes about equally effectively. Thus, unlocalized amplitude information, including the dominant orientation of an image, appears to be of limited use for recognizing scene gist.