Purchase this article with an account.
Robert J. Peters, Laurent Itti; A computational model of task-dependent influences on eye position. Journal of Vision 2006;6(6):512. doi: https://doi.org/10.1167/6.6.512.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Computational models of bottom-up attention can perform significantly above chance at predicting eye positions of observers passively viewing static or dynamic images. Nevertheless, much of eye movement behavior (50% or more) is unexplained by purely bottom-up models, and is typically attributed to top-down, inter-observer, task-dependent, or random effects. Other studies have qualitatively described such high-level effects in naturalistic interactive visual tasks (e.g., while driving, how often do people fixate other cars, or the road, or road signs); yet the underlying neurocomputational mechanisms remain unknown. Here, we introduce a simple computational model of task-related eye position influences in interactive tasks with dynamic stimuli. This model extracts from each frame a low-dimensional feature signature (“gist”), compares that with a database of eye position training frames, and produces an eye position prediction map. Finally, we combine the task-related and bottom-up maps, and compare the combined maps with observers' actual eye positions across 216000 frames from 24 five-minute videogame-playing sessions. For analysis, each map was rescaled to have zero mean and unit standard deviation; the average predicted value at human eye position locations was 0.61 ± 0.1 in the purely bottom-up maps, and 2.42 ± 0.07 in the combined maps (a random model gives an average value of 0). Thus, this straightforward model of task-dependent effects offers some of the strongest purely computational general-purpose eye movement predictions to date, going significantly beyond what is explained by purely bottom-up effects; yet it relies only on simple visual features, without requiring any high-level semantic scene description.
This PDF is available to Subscribers Only