A structural description model of entry- and instancelevel object classification

Purpose People can easily classify objects at multiple levels of specificity. For example, a given chair might be classified as "chair" (entry-level), "office chair" (subordinate-level) or "my office chair" (instance-level). Most models of human object recognition account for recognition at only one level of specificity-either entry-level or subordinate-/instance-level classification. For example, most structural description models can classify objects as a member of a specific entry-level class, however these same models cannot classify an object into their subordinate- or instance-level class. By contrast, most view-based models can easily recognize an object as a specific instance, but they do not provide a natural account of entry-level classification (e.g., Poggio & Edelman, 1990). Method. We will describe a model that represents object shape in a form that supports recognition at multiple levels of specificity. Following Biederman (1987; Hummel & Biederman, 1992), the model represents objects in terms of their parts and their inter-relations. But unlike models based on strictly categorical structural descriptions, the current model represents shape in a nonlinear fashion that emphasizes categorical attributes without discarding all metric information. Used as the input to object-selective units with Gaussian receptive fields, the resulting representation of shape provides a natural basis for recognizing objects at multiple levels of specificity: Units with narrow (small a) receptive fields recognize objects at the instance level, while units with large o recognize objects at the entry level, Results, Simulations results show that the model accounts for a number of phenomena, including our ability to recognize objects at multiple levels of specificity, our ability to recognize novel instances of known object classes, and the finding that entry-level classification occurs before subordinate-level classification. The model also implicates attention in shape perception, and predicts that visual attention is more important for instance- than entry-level recognition. Conclusions. Within a single representation of object shape, it is possible to account for how humans classify an object as a specific instance, along with classifying it as a member of a general class. NONE.