Perceiving animacy from shape

Abstract

Superordinate visual classification for example, identifying an image as animal, plant, or mineral is computationally challenging because radically different items (e.g., octopus, dog ) must be grouped into a common class ( animal ). It is plausible that learning superordinate categories teaches us not only the membership of particular (familiar) items, but also general features that are shared across class members, aiding us in classifying novel (unfamiliar) items. Here, we investigated visual shape features associated with animate and inanimate classes. One group of participants viewed images of 75 unfamiliar and atypical items and provided separate ratings of how much each image looked like an animal, plant, and mineral. Results show systematic tradeoffs between the ratings, indicating a class-like organization of items. A second group rated each image in terms of 22 midlevel shape features (e.g., symmetrical, curved ). The results confirm that superordinate classes are associated with particular shape features (e.g., animals generally have high symmetry ratings). Moreover, linear discriminant analysis based on the 22-D feature vectors predicts the perceived classes approximately as well as the ground truth classification. This suggests that a generic set of midlevel visual shape features forms the basis for superordinate classification of novel objects along the animacy continuum.