Two roles for attention in shape perception: A structural description model of visual scrutiny

This paper presents MetriCat, a model of the human capacity to recognize objects both as members of a general class (e.g. "chair") and as specific instances ("my office chair"), and of the role of visual attention in this capacity. MetriCat represents the attributes of an object's parts and their relations in a non-linear fashion that provides a natural basis for recognition at both the class and instance levels (Stankiewicz & Hummel, 1996). Like previous structural description models (e.g. Hummel & Biederman, 1992), MetriCat represents part attributes and relations independently, dynamically binding them into structural descriptions. The resulting representation suggests two roles for visual attention in shape recognition: attention for binding and attention for signal-to-noise control. MetriCat implements both functions as special cases of a single mechanism for controlling the synchrony relations among units representing separate object parts. The model accounts for the time course of class- and instance-level classification, and makes several predictions about the relationships between attention, time and levels of classification.