Academic journal article Memory & Cognition

View Combination in Scene Recognition

Academic journal article Memory & Cognition

View Combination in Scene Recognition

Article excerpt

Becoming familiar with an environment requires the ability to integrate spatial information from different views. We provide evidence that view combination, a mechanism believed to underlie the ability to recognize novel views of familiar objects, is also used to recognize coherent, real-world scenes. In two experiments, we trained participants to recognize a real-world scene from two perspectives. When the angular difference between the learned views was relatively small, the participants subsequently recognized novel views from locations between the learned views about as well as they recognized the learned views and better than novel views situated outside of the shortest distance between the learned views. In contrast, with large angles between training views, all the novel views were recognized less well than the trained views. These results extend the view combination approach to scenes and are difficult to reconcile with models proposing that scenes are recognized by transforming them to match only the nearest stored view.

As mobile organisms, humans benefit from the ability to recognize places and scenes from vantage points that they have not previously experienced. This ability implicates the existence of psychological mechanisms that compare spatial information from current and previously experienced perspectives. Although the existence of these mechanisms is self-evident, theoretical formulations of how they enable people to recognize scenes have yet to be fully developed, and empirical evidence that would constrain these theories is sparse. In contrast, models of the processes that underlie people's recognition of single objects have undergone significant theoretical development (Biederman, 1987; Biederman & Gerhardstein, 1993; Bülthoff & Edelman, 1992; Bülthoff, Edelman, & Tarr, 1995; Edelman, 1999; Tarr & Pinker, 1989; Ullman, 1989, 1996), and a wealth of results bears on the question of how prior views of objects enable recognition of novel views.

According to a normalization account of object recognition, people store a number of exemplar views of objects; to recognize a novel view, they transform the novel percept so that it matches the nearest (i.e., most similar) stored exemplar of the object (Tarr, 1995; Tarr & Pinker, 1989; Ullman, 1989). Normalization models assume that even when a novel view is equidistant from two (or more) stored representations, it is normalized with respect to only one of them. Thus, one of the behavioral implications of a normalizing process is that a graceful, roughly monotonie increase in error rate, recognition time, or both will be observed as a function of the distance between the novel view to be recognized and the nearest learned view. It is in this sense that a normalization approach predicts that recognition will be viewpoint dependent.

An alternative account of object recognition, known as view combination, maintains that the extent to which a novel view of an object can be readily recognized depends on its degree of structural similarity to a set of multiple stored views (Bülthoff & Edelman, 1992; Edelman, 1999; Edelman & Bülthoff, 1992; Edelman, Bülthoff, & Bülthoff, 1999). In contrast to the degradation in performance predicted if recognition occurs by normalization, the view combination approach predicts that it is possible for some novel views to be recognized as well as familiar views, because information from two or more structurally similar stored views can be combined to facilitate recognition. However, recognition should not be facilitated if the view to be recognized is too structurally disparate from the familiar (stored) views.

More formally, the predictions of the view combination approach arise because objects are represented as points in a multidimensional shape space that is spanned by their parametric similarities to a small number of reference objects, which may be construed as prototypes (Edelman, 1999). …

Search by... Author
Show... All Results Primary Sources Peer-reviewed

Oops!

An unknown error has occurred. Please click the button below to reload the page. If the problem persists, please try again in a little while.