Abstract
How is scene gist recognized? We investigated the roles of the amplitude spectra and local relative phase of scenes in both scene gist recognition and masking. We hypothesized that the information most useful for scene gist recognition also produces the strongest scene gist masking. Previous results (Loschky, et al., Psychonomics, 2005) showed that masks having the amplitude spectra of scenes but random phase spectra produced stronger masking than white noise masks, but weaker masking than scene image masks. We tested the further hypothesis that masks sharing both similar amplitude spectra to scenes and similar local relative phase information would produce stronger scene gist masking than masks sharing only the amplitude spectra of scenes. We compared scene gist masking produced by 1) synthesized textures (Portilla & Simoncelli, 2000) based on scenes, and 2) phase-randomized scenes using the RISE algorithm (Sadr & Sinha, 2004). Results confirmed our hypothesis.
Scene categories also varied in necessary and sufficient information for gist. Similar patterns were found for both the gist recognition threshold of phase randomization, and for synthesized texture images: forests and mountains were the most recognizable. Furthermore, compared to masking by an image from a different scene category, scenes masked by either their fully phase-randomized versions or their synthesized texture versions showed the same across-category pattern: forests and mountains were more recognizable than other categories under conditions of target/mask integration. Thus, some categories are more recognizable based on amplitude and local relative phase information, while others require more information (e.g., global relative phase—‘layout’).
SR Research: provision of Experiment Builder software & technical support