Abstract
When viewing a scene, humans rapidly move their eyes to foveate visual features and objects of interest. In natural conditions, this process is temporally complex, yet little is known about how the temporal structure of naturalistic stimuli affects the dynamics of eye movements under free viewing. We tracked eye position while observers watched a 6-minute scene from a feature film that was shot as a continuous sequence (with no cuts). Consistent with previous reports (Hasson et al., J Neurosci, 2008), eye movements were highly reliable, both across repeated presentations and across observers. We then divided the scene into clips of various durations (ranging from 500 ms to 30 s) and scrambled the temporal order of the clips, thereby introducing cuts. Eye-movement reliability, quantified as the covariance between eye positions to the scrambled clips and those during the corresponding portions of the full-length scene, was found to increase as a power-law function of clip duration, from ∼0 for the 500 ms clips to an asymptote for clips >30 s in duration. We developed a model that assumed that observers searched randomly following each cut, fixating at arbitrary locations until finding a target of interest and then tracking it faithfully. We fit the model to the data by analytically deriving the model's prediction for the relationship between clip duration and eye movement reliability (covariance). While simple, this model fit the data well with only two free parameters (number of possible target locations, asymptotic covariance). However, the model fits exhibited a systematic bias at the shortest scramble durations. We conclude that exploratory fixations depend critically on the temporal continuity of stimuli, and that human observers might utilize a random search strategy when viewing naturalistic, time-varying stimuli.