Purchase this article with an account.
Barbara Hidalgo-Sotelo, Aude Oliva; History repeats itself: A role for observer-dependent scene context in visual search. Journal of Vision 2010;10(7):1304. doi: https://doi.org/10.1167/10.7.1304.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Eye guidance during visual search and naturalistic scene exploration is based on combining information from image-based cues and top-down knowledge (e.g. target features, Zelinsky, 2008; scene context region, Torralba et al, 2006). It is not known whether previous searches of a scene contribute to search guidance. How much information is gained by knowing an observer's history of searching familiar scenes? To probe this question, we recorded eye movements while observers searched for a camouflaged book in indoor scenes (100% target prevalence). In the repeated condition, scenes were searched 8 times by the same observer, while in the novel condition each scene was searched once. This large dataset of search fixations was used to evaluate the similarity between scene locations fixated during an observer's repeated search relative to novel searches of the same scene.
An ROC analysis was used to evaluate how accurately fixated locations were predicted by distributions representing several types of top-down knowledge: (1) Categorical scene context: fixations drawn from different observer's search of a novel scene, (2) Learned scene context: fixations drawn from different observer's repeated searches of the scene; and (3) Observer-dependent scene context: fixations from one observer's repeated searches of the scene. The results reported below used the first three search fixations of each trial, but similar results were obtained using the first fixation exclusively. Categorical scene context predicted fixated locations of different, novel searchers with a high degree of accuracy (84%). Learned scene context, based on different searcher's repeated fixations, was similarly accurate (85%). Observer-dependent scene context, interestingly, provided a significant improvement in prediction accuracy relative to baseline controls and other forms of context (90%). In summary, having an observer's history of search fixations in a specific scene provides, on average, more accurate and less variable predictions of where that observer is likely to look.
This PDF is available to Subscribers Only