Abstract
What makes two shapes similar? Given two shapes, is there a mathematically principled way to predict human similarity judgments as they’re used to classify both 2D and 3D shapes? In previous work, we proposed a shape similarity measure based on shape skeletons in a Bayesian framework. This measure posits that shape similarity is a function of the probability that the two shapes were generated from a common skeletal model. A key term in this probability is a probabilistic cost of transforming one shape into the other using the skeleton-based shape generation process. In current work, we expand on our previous validation of this model, comparing it to other successful models in a pair of shape classification experiments. We then use the model to test whether humans tend to infer prototype-like process (which estimates a common model from multiple examples) or exemplar-like process (which does not). We ran two experiments to test human shape categorization. Exp. 1 uses 2D shape silhouettes, while Exp. 2 uses images of 3D shapes. In both experiments, subjects are shown examples from a novel category of unfamiliar shapes, and (in some conditions) a negative example of a shape not from that category. Subjects are then asked to select shapes from a 6×6 grid of shapes that they judge as belonging to the novel category. We then use our model’s similarity metric to predict the classification judgments from the experiments, and compare our model’s performance to that of deep learning models on the same data. Finally, we use two variations on our model, in conjunction with the human classification patterns, to test whether humans infer prototype-like or exemplar-like shape category structures. Here, we find that our prototype-like model provides a better fit to human data than does our exemplar-like model.
Acknowledgement: NIH (NEI) Grant EY021494