Abstract

R. J. Peters, L. Itti, Integrating low-level and high-level visual influences on eye movement behavior, In: Proc. Vision Science Society Annual Meeting (VSS07), May 2007. (Cited by 1)

Abstract: We propose a comprehensive computational framework unifying previous qualitative studies of high-level cognitive influences on eye movements with quantitative studies demonstrating the influence of low-level factors such as saliency. In this framework, a top-level ''governor'' uses high-level task information to determine how best to combine low-level saliency and gist-based task-relevance maps into a single eye-movement priority map. We recorded the eye movements of six trained subjects playing 18 different sessions of first-person perspective video games (car racing, flight combat, and ''first-person shooter'') and simultaneously recorded the game's video frames, giving about 18 hours of recording for 15,000,000 eye movement samples (240Hz) and 1.1TB of video data (640x480 pixels at 30Hz). We then computed measures of how well the individual saliency and task-relevance maps predicted observers' eye positions in each frame, and probed for the role of the governor in relationships between high-level task information -- such as altimeter and damage meter settings, or the presence/absence of a target -- and the predictive strength of the maps. One such relationship occurred in the flight combat game. In this game, observers are actively task-driven while tracking enemy planes, ignoring bottom-up saliency in favor of task-relevant items like the radar screen; then, after firing a missile, observers become passively stimulus-driven while awaiting visual confirmation of the missile hit. We confirmed this quantitatively by analyzing the correspondence between saliency and eye position across a window of +/-10s relative to the time of 328 such missile hits. Around -200ms (before the hit), the saliency correspondence begins to rise, reaching a peak at +100ms (after the hit) of 10-fold above the previous baseline, then is suppressed below baseline at +800ms, and rebounds back to baseline at +2000ms. Thus, one mechanism by which high-level cognitive information can influence eye movements is through dynamically weighting competing saliency and task-relevance maps.

Themes: Computational Modeling, Model of Bottom-Up Saliency-Based Visual Attention, Model of Top-Down Attentional Modulation, Human Eye-Tracking Research