Abstract
Some objects look more similar than others on the basis of their shape. The current state-of-the-art models of object shape come from hidden layers of deep convolutional neural nets; however, the parameterization of shape is embedded in millions of parameters. Here, we consider a recently developed alternative modelānormalized contour curvature (NCC), as a simpler model to quantify shape structure. This model takes an isolated object image and computes a histogram of curvature-radius values over the level sets of the pixel intensities. We first characterized the overall structure of shape space captured by the NCC model, by computing NCC features (n=51) over a dataset of isolated inanimate objects (N=7000), and employing a principle component analysis. The first four principal components (PCs) explained more variance than a chance reshuffling of the parameters. We next tested the behavioral relevance of these four shape axes. In Experiment 1, behavioral measures of curviness scores were obtained on a test set of 72 images (N=28), using a 5-point curvy-to-boxy scale (Long et al., 2017). In Experiment 2, we obtained measurements of overall shape similarity by having observers (N=16) arrange 72 items in a circular arena so that similar shaped things were nearby. Behavioral ratings of curvature were strongly associated with the 2nd PC score (r=0.64), indicating that perceptual shape dimension is relatively well isolated by this axis of NCC shape space. Additionally, overall shape similarity judgments were modeled within the noise ceiling, placing high weights on the first three principle components (š=0.27, noise ceiling: 0.25-0.33). Taken together, this work demonstrates that the normalized contour curvature model can summarize shape space with a relatively small number of parameters, where the major axes through this shape space are meaningfully related to perceived shape and curvature of inanimate objects.
Meeting abstract presented at VSS 2018