A relationship between human shape categorization and the statistics of natural shapes

We studied the classification of shapes into broad natural categories, such as"animal" and "leaf", into which many shapes can proceed without overt basic-level recognition. Many shape representation models make implicit assumptions about what shape structures often occur in natural shapes, but such assumptions are not generally closely tied to real-world measurements. In order to tune a model of shape classification to the natural environment we collected statistics from a large database of real animal and leaf shapes; first we calculated the MAP skeletons of these shapes, and then computed several different statistical properties of the skeletons. These statistics allow for the creation of an "ecologically-informed" shape classification model that can generalize of many of the specific structures observed in the two classes. To investigate human shape classification subjects were shown shapes created by taking a weighted average of an animal and a leaf shape, resulting in a novel morphed shape. The task was to classify a shape as animal or leaf. Subjects easily performed this task, and their responses were strongly related to the weight used to averaged the shapes. Next, the classifier was used to obtain the likelihoods that a given morphed shape belonged the each class. These likelihood values are predictive of the subjects' responses, suggesting a relationship between human classification of shapes and the shapes' skeletal structures.