The function of VLTM in scene perception

In what manner does VLTM for a scene influence perceptual processing of that scene? First of all, VLTM for scenes allows us to recognize scenes and categorize them. However, there has been surprisingly little research examining the mechanisms of scene identification. Initial evidence suggests that scene identification depends on global scene properties rather than local analysis of constituent objects (Oliva & Torralba, 2006). In addition, scene identification is extraordinarily rapid (Potter

& Levy, 1969; Thorpe, Fize, & Marlot, 1996). Efficient scene identification raises the possibility that scene memory might influence even fairly early perceptual operations over a scene. I shall first consider whether scene identification influences the perceptual recognition of objects in a scene. I then examine the role of scene knowledge in guiding attention to task-relevant areas of a scene.

Effects of scene memory on object recognition

Hollingworth and Henderson (1998) identified three possible means by which one’s knowledge about a particular scene type (e.g., that kitchens tend to contain stoves but not motorcycles) could influence the identification of constituent objects. First, scene knowledge could interact with early visual processing to enhance the perceptual description of scene-consistent objects (description enhancement). Second, scene knowledge could influence the comparison of perceptual object representations to stored category representations, lowering goodness-of-fit thresholds for consistent object categories (criterion modulation). Third, scene knowledge could be isolated from object recognition operations, influencing only postperceptual reasoning (functional isolation).

When examining the influence of scene knowledge on the perceptual recognition of objects, it is critical to ensure that participants cannot use their knowledge of scenes to make an educated guess. For example, if one is blindfolded, taken into a kitchen, and asked to name the large appliance in the corner, one could reason that the probed object is likely to be a stove or a refrigerator (rather than a washing machine or an air conditioner) in the absence of any visual input at all. Early studies examining the effects of scene context on object recognition found that semantic- ally consistent objects (e.g., a computer in an office) were recognized more accurately than inconsistent objects (e.g., a computer in a bathroom) (Biederman et al., 1982; Palmer, 1975). However, educated guessing was not adequately controlled in these studies.The consistent-object advantage could have derived from the fact that participants were biased to report consistent objects, without any direct effect of scene context on the perceptual mechanisms of object recognition.

To provide a better measure of scene context effects on perceptual object recognition, Hollingworth and Henderson (1998) used a 2-AFC method similar to that developed by Reicher (1969; see also D. D. Wheeler, 1970) to examine the effects of word context on letter identification. On each trial, participants saw a brief display of a scene containing either a semantically consistent target object or an inconsistent target object. The scene was followed by two object labels of equivalent consistency. For example, a kitchen scene (or, in the inconsistent condition, a farm scene) contained a mixer target object followed by the labels “mixer” and “coffee maker”. Because the two alternatives were both either consistent or inconsistent with the scene, educated guessing on the basis of scene knowledge could not influence performance. With this control over guessing, no advantage for the detection of consistent objects was observed, supporting the functional isolation hypothesis. The Hollingworth and Henderson (1998) results indicate that we accurately see what is present in a scene and not necessarily what we expect to see. Given the opportunity to guess, however, biases generated by scene knowledge will influence report.

Recently, Davenport and Potter (2004; see also Davenport, 2007) revisited the issue of scene context effects. In their paradigm, participants viewed stimuli consisting of a background scene and a prominent foreground object, with the consistency between the two manipulated. After brief presentation of each scene, participants named the foreground object. Davenport and Potter observed more accurate naming of consistent versus inconsistent objects. However, these experiments represent something of a methodological step backward, because Davenport and Potter did not adequately control educated guessing. In this naming paradigm, when an object was not fully identified, participants could use their knowledge of the scene to bias the naming response toward consistent objects (see Palmer, 1975), as the target was more likely to be one of the relatively small set of objects consistent with the scene than one of the large set of objects inconsistent with the scene. Davenport and Potter did include a guessing correction that involved subtracting incorrect reports of consistent objects from correct reports, but simple subtraction is not sufficient when bias could be influencing report (Green & Swets, 1966). In general, any paradigm with an unbound set of alternatives (as in naming) is subject to selection biases that can be very difficult to eliminate. It was precisely for this reason that Reicher (1969) developed the 2-AFC method used by Hollingworth and Henderson (1998).

In summary, current evidence indicates that when educated guessing is adequately controlled, consistent objects are detected no more efficiently than are inconsistent objects. This does not imply, however, that there are no effects of scene knowledge on the perceptual processing of objects. Scene knowledge can guide attention to particular objects in a scene that are relevant to the current task, reviewed below. In addition, context influences the extent of perceptual and cognitive processing devoted to an object. For example, inconsistent objects, once identified, are fixated longer in a scene than are consistent objects (Henderson, Weeks, & Hollingworth, 1999).

< Prev   CONTENTS   Source   Next >