Scene Structure and Change Detection Running Head: Scene Structure and Change Detection Scene-like Structure Enhances Change Detection

Does scene-like structure enhance change detection? Theories of scene perception, neural representation of scenes and object recognition seem to imply that it should. However, extant evidence from the change detection literature is thus far ambiguous with respect to this question. The four experiments reported here demonstrate that scene-like structure does, in fact, enhance change detection. In Experiments 1, 2 and 4, change detection was better for normal scenes than scenes whose scene-like structure had been destroyed by jumbling. In Experiment 3, inversion failed to interfere with change detection, demonstrating that the disruption of surface and object continuity inherent to jumbling is responsible for reduced change detection. These findings provide a crucial commonality between change detection research, and theories of perception and neural representation that suggest structure in the world constrains how we see. 2 Scene Structure and Change Detection Theorists have long noted that the visual world contains many structural regularities that the perceptual system can use to recognize objects and interact with the environment (e.g. Biederman, 1987; Gibson, 1986). Recent work has focused on the visual system’s tendency to rely on regularities in the world as a sort of “external memory” (O'Regan, 1992; O'Regan & Noe, 2001) to account for phenomena such as change blindness (Rensink, 2000, 2002). Other research has demonstrated that the visual system can take advantage of regularities in a scene to guide attention to the location of objects of interest (Chun & Jiang, 1998; Chun & Nakayama, 2000) and detect changes to objects that have a high probability of changing (Beck, Angelone, & Levin, 2004). In the current report, we tested whether detecting visual changes in well-structured scenes is easier than detecting changes in jumbled (i.e. unstructured) scenes, because wellstructured scenes possess many of the regularities that the visual system typically relies upon. Before describing research exploring the impact of scene structure on object perception, we would like to briefly specify what we mean by scene structure. Our discussion will focus on real-world scenes, either in natural or in-door environments, and we will assume that scenes are characterized by a clear spatial structure that can support an understanding of the relative location of objects in the scene. The most important parts of this structure are the (usually) immobile surfaces that support objects, confine locomotion, and provide cues for navigation were one actually in the scene. This definition closely follows that given by Henderson and Hollingworth (1999). A critical point to emphasize is that their definition includes reference to the semantic coherence of the scene, and that our definition of scene structure may also include some global 3 Scene Structure and Change Detection semantic identity of the scene such as gist. However, we do not include the meaning of objects in the scene, or information about interobject relationships and jointly specified functions. This restriction stems primarily from recent findings that scene-responsive brain regions (e.g. the PPA or parahippocampal place area) respond minimally to semantic qualities of the scene such as familiarity (Epstein, 2005). Thus, we focus primarily on how the perceptual structure of a scene might affect change detection. Does scene-like structure facilitate change detection? Various lines of research are consistent with the idea that scene organization affects how a scene is perceived and represented, and might therefore be expected to affect change detection. For example, people can identify objects more accurately in coherently structured scenes than jumbled scenes (Biederman, 1972) and the abovementioned PPA seems dedicated to representing the local geometry of structured scenes in both perceptual and visual imagery tasks (Epstein & Kanwisher, 1998; O'Craven & Kanwisher, 2000). More recent research has demonstrated that observers take longer to determine spatial relationships of locations that straddle local discontinuities in scenestructure than locations that do not straddle such discontinuities (Sanocki, Michelet, Sellers, & Reynolds, 2006). All together, this research implies that scene-structure should affect change detection because change detection relies in part on object recognition, neural representation and computation of spatial relationships. However, the tasks used in this previous research (e.g. identifying objects, simply viewing scenes, and determining spatial relationships) do not require all of the representational and comparison processes necessary to detect changes across views (Simons, 2000). Thus, in 4 Scene Structure and Change Detection principle this research could be consistent with the idea that scene structure does not affect change detection. Moreover, research using change detection methods is also equivocal. Some research suggests that Gestalt principles influence change detection; cues that direct attention to a particular object confer an advantage in change detection for the cued object and objects that are part of the cued object’s Gestalt group (Woodman, Vecera, & Luck, 2003). However, this study did not use realistic scenes as stimuli, leaving open the possibility that the organizing perceptual structures characteristic of natural scenes are less effective in facilitating change detection. Furthermore, various studies of CB using the flicker paradigm (Rensink, O'Regan, & Clark, 1997) and realistic scenes suggest that high-level semantic factors are important determinants of the ease of change detection, whereas scene organization is, by itself, relatively unimportant. In the flicker paradigm, observers view cyclically alternating versions of a scene separated by a brief blank screen or ‘flicker’. The time it takes observers to find the difference between the images is the dependent variable. Several semantic factors are capable of attenuating CB. For example, objects that are of central interest (i.e. highly relevant) to the meaning of a scene are typically detected faster than objects of marginal interest (O'Regan, Rensink, & Clark, 1999; Rensink et al., 1997), as are objects that are semantically inconsistent with the scene’s context (Hollingworth & Henderson, 2000). Further, inverting a scene does not influence the time it takes to find a change, but inversion of a scene can attenuate the central interest advantage in some cases (Kelley, Chun, & Chua, 2003; Shore & Klein, 2000). Accordingly, scene inversion may affect the ease with which a scene’s meaning can be accessed, and can influence change detection. 5 Scene Structure and Change Detection However, disrupting scene-like structure does not seem to have very strong effects on change detection, at least in the flicker paradigm (Yokosawa & Mitsumatsu, 2003). In experiments by Yokosawa and Mitsumatsu, participants searched for changes in scenes that were divided into 24-grid sections. The experimenters manipulated scene structure in two ways: 1) by removing some of the grid sections and 2) by jumbling the grid sections (i.e. moving grid sections out of their original positions). Removal of grid sections had a large effect; observers generally took longer to find changes when there were more visible grid sections. This increase in RT is not very surprising, because adding grid sections increases the area that must be searched. The interesting finding is that under most conditions, jumbling a scene had virtually no effect on change detection. One exception was that changes were detected more efficiently in normal scenes, but only when comparing completed and nearly completed scenes, and only when observers knew that scenes would always be normal (as opposed to jumbled). Basically, detecting a change in a 70%-complete normal scene took as long as detecting the change in a whole scene; but it took less time to detect a change in a 70%-complete jumbled scene compared to a 100%-complete jumbled scene. The results from Yokosawa and Mitsumatsu (2003) appear to confirm that a coherent scene structure can facilitate change detection, in at least one situation. However, the small jumbling effect they found is difficult to interpret, because it is possible that observers searched near-complete normal scenes as if they were complete scenes. If this were the case then the ‘improvement’ in efficiency would not be due to a facilitative effect of scene-like organization on change detection, but to disruptive effect of coherent scene organization for near-complete scenes. This possibility and the fact 6 Scene Structure and Change Detection that jumbling a scene had no effect on change detection in other conditions (even other conditions comparing 70% and 100% complete scene; see Experiment 3 in Yokosawa & Mitsumatsu, 2003), conflict with the idea that scene structure per se can influence change detection. Thus, evidence from the scene perception, cognitive neuroscience and change detection literatures suggest that jumbling a scene should affect change detection, but the idea has not yet been strongly supported. Given the significance of organizational structure for theories of object recognition, scene perception and cognitive neuroscience, it is important to determine whether scene-like structure can enhance change detection, because change detection tasks rely in part upon processes that these theories purport to explain. If change detection were not affected by scene structure, it would suggest that mechanisms for detecting change might be independent of mechanisms involved in scene perception. This would indeed be surprising, but since there is no good evidence that scene-like structure enhances change detection, and some evidence that scene-like structure does not affect change detection, it remains possible.