Differing views on views: comments on Biederman and Bar (1999)

In a recent article, Biederman and Bar (1999) present several results to support ‘‘a class of theories [that] assumes that non-accidental properties (NAPS) might be exploited so that even novel objects can be recognized under depth rotation’’ — specifically, theories based on ‘geons’ (Biederman, 1987; Hummel & Biederman, 1992). Biederman and Bar likewise present results that they believe to be inconsistent with a ‘‘class of theories ... based on generalization from templates specified by metric properties’’ — specifically, ‘viewbased’ or ‘image-based’ theories (Bricolo, Poggio, & Logothetis, 1997; Poggio & Edelman, 1990; Tarr & Bülthoff, 1995, 1998). Because our disagreements with Biederman’s theoretical approach have been detailed in many other forums (Hayward & Tarr, 1997; Tarr & Bülthoff, 1995, 1998) we do not feel it is necessary to respond in kind. We do feel, however, that it is necessary to address a series of straightforwardly incorrect claims made by Biederman and Bar regarding our published results. Below we enumerate these claims, and our replies to each: 1. ‘‘A task in which subjects are trained with arbitrary names for a particular pose of an object (Tarr, Williams, Hayward, & Gauthier, 1998, Experiment 3; Haywood (sic) & Tarr, 1997, Experiment 2), particularly if the distinguishing information is difficult to discriminate, is problematic insofar as pose is part of what is learned and, potentially, used’’ (p. 2895). According to this statement, our cited experiments produced results that are artifactual because names for objects were learned only at specific viewpoints. There are at least three reasons to conclude that no such problems exist. First, in these and other studies, we obtained the same pattern of viewpoint dependency with identification (naming) tasks and with Biederman and Bar’s preferred same–different matching tasks (Hayward & Tarr, 1997; Tarr, Bülthoff, Zabinski, & Blanz, 1997; Tarr et al., 1998). Moreover, we almost always employed a same–different matching procedure that was identical to (and intentionally based on) that used by Biederman and Gerhardstein (1993). Second, in these same three studies (Hayward & Tarr, 1997; Tarr et al., 1997; Tarr et al., 1998) we explicitly used stimuli (‘geons’ and ‘geon objects’) which were extremely easy to discriminate. Indeed, most stimuli were almost identical to those used by Biederman and Gerhardstein (1993). These studies show the same pattern of responses as studies using highly similar objects that are difficult to discriminate (e.g. Tarr, 1995; Tarr & Pinker, 1989). Thus, there is relatively little evidence, as suggested by Biederman and Bar, that recognition of highly similar objects is qualitatively different from recognition of geons (for additional evidence, see Hayward & Williams, in press). Third, Biederman and Bar appear to be intimating that pose should be divorced from training on novel objects. To the extent that this is possible, it has certainly been accomplished in several studies, either by presenting animated objects rotating back and forth in depth (Bülthoff & Edelman, 1992; Edelman & Bülthoff, 1992), by training subjects with the standard and mirror-reversed versions of each object (so that subjects would never have to do the equivalent of distinguishing a p from a q or a b from a d) (Tarr, 1995; Tarr & Pinker, 1989), or by training subjects with multiple views of each object (Tarr, 1995; Tarr & Gauthier, 1998; Tarr & Pinker, 1989). Such manipulations would * Corresponding author. Tel.: +852-26096195; fax: +85226035019. E-mail addresses: [email protected] (W.G. Hayward), michael–[email protected] (M.J. Tarr). 1 Comments can also be addressed to MJT. Tel.: +1-401-8631148; fax: +1-401-8632255. 2 Because some of the points raised in this letter relate directly to the appearance of stimuli we have used in various experiments, readers are encouraged to examine our stimuli for themselves — all of which are available for download at our web site: http:// www.cog.brown.edu/ tarr.