Eye movements during a change detection search task

Purpose. Specify factors affecting our ability to detect changes in realistic scenes using a "flicker" search paradigm (Rensink et. al., ARVO 96). Methods. Scenes depicted color images of objects (tools or toys) arranged on an appropriate surface (workbench or crib). Subjects viewed a repeating sequence of three displays (scenel → gray → scene2 → gray...) and indicated the presence or absence of a changing object between scenes 1 and 2 (these scenes were identical except for a single object in Vz of the trials). Manipulations included: 1) change vs. no-change, 2) set size (1, 3, or 9 items), 3) orientation of changing objects (same vs. rotated 90°), and 4) within- vs. across-category change (e.g., a hammer in a workbench scene changing with either a wrench or a doll). Results. Object changes violating orientation or category were detected faster and fixated sooner relative to the no-change baseline. Importantly, these effects also interacted with set size, with changes in the 9 item displays showing the greatest benefit. Eye movement data indicated the expression of these interactions very early in search, often as soon as the second saccade. Conclusions. One reason why earlier studies using realistic scenes have found such poor change detection performance is that the stimuli used contained large numbers of objects. When this factor is controlled, change detection varies with set size and visual conspicuity like many other search tasks, with object realism or embeddedness in a scene having little effect. Preferential fixation of changing objects also implies a comparison between object properties stored in memory and those properties obtained from the visual periphery.