Differing views on views: response to Hayward and Tarr (2000)

Abstract

At the outset we should emphasize that nothing in Hayward and Tarr’s commentary speaks to the main results of the Biederman and Bar (1999) report: Slight costs of rotation in detecting geon differences when matching a sequential pair of novel objects (3.3% increase in error rates), but massive costs (46.2% increase in error rates) in the detection of metric differences. Their comments are all addressed to our review of the literature of the relatively small costs (versus zero costs) that have been observed when geon differences were, presumably, available. Although they characterize some of Biederman and Bar’s arguments as ‘incorrect’, in fact these arguments can readily be defended. Given the sensitivity that prompted these authors to write their comments, it is disappointing that their own characterization of many of Biederman and Bar’s points are inaccurate. It is not completely clear why the origin of these relatively small rotation costs are an important theoretical issue to Hayward and Tarr insofar as they have now ‘eschewed’ mental rotation mechanisms (their note 2). Instead, viewpoint costs are now regarded by these authors as reflecting changes in information and that ‘there is no directly causal relation between the magnitude of a change in viewpoint of an object, and the magnitude of the associated cost in recognition’ (note 2). This has been our position all along (Biederman, 1987; Biederman & Gerhardstein, 1993, 1995; Biederman & Bar, 1999). 1. Why assigning arbitrary names to novel objects may be problematic for assessing the view dependence/ independence of object representations Consider a visual classification task in which brief masked pictures of tables are presented to the left of fixation and brief masked pictures of chairs are presented to the right of fixation. An observer required to distinguish between chairs and tables might readily use the left–right view information to achieve high accuracy on such a task. However, few would accept such evidence of ‘view-based recognition’ as addressing issues concerned with the ‘mental representation of objects’ especially if that mental representation was to be interpreted as a representation of shape. An alternative position, proposed by Biederman and Cooper (1992), is that these viewpoint effects may be part of the episodic representation that specifies view variables and situation variables bound to a representation, perhaps implicit, of shape. That there might be at least two representations of objects is evidenced by Biederman and Cooper’s (1992) finding that changes in the size of an image had no effect on basic-level name priming but produced marked interference in old-new, episodic recognition judgments. It is important to note that the first block of trials — basic level naming — was identical for the two tasks (Unlike experimental trained naming of novel stimuli, basic level naming is well practiced to be independent of view variables, such as an object’s position in the visual field or whether it is facing left or right). Biederman and Cooper’s interpretation was that the basic-level name priming was mediated by a representation invariant to view variables but that the old–new judgments were mediated by an episodic representation that combined shape and viewvariables. Similar results (no effects on priming but large costs on old judgements) were reported by Biederman and Cooper (1991a,b) and Cooper, Biederman, and Hummel (1992) for position, reflection, and orientation. Hayward and Tarr’s comments would imply that there is a single, all-purpose, representation and that any variable that affects object recognition performance is, ipso facto, reflecting this single representation. Our position is perhaps more consistent with the classical * Corresponding author. Tel.: +1-213-7406094; fax: +1-2137405687. E-mail address: bieder@usc.edu (I. Biederman).