Automated Generation of Perceptually-Uniform Circular Spaces for Novel Naturalistic Shapes

Shape processing holds a crucial role in higher-level visual functions, such as object recognition. Previous efforts to explore high-level vision, independent of shape influence, involved meticulous measurements to control for perceptual shape similarity across distinct stimulus sets. However, there is a compelling need for a more efficient approach to automatically synthesize perceptually uniform spaces of novel shapes. In this pursuit, we present an image-based method that autonomously generates numerous perceptually uniform and circular shape sets, eliminating the need for extensive psychophysical measurements. Our method employs a search for circular shape sets correlated with ShapeComp, an image-computable shape similarity metric based on over 100 descriptors, highly predictive of human shape similarity. Using multi-arrangement methods, we demonstrate that predicted human similarity arrangements of shape sets defined as circular and uniformly spaced by ShapeComp, align with human shape similarity judgments and approximate circularity. Notably, shape sets chosen for uniformity and circularity in alternative shape spaces (e.g., Generative Adversarial Networks or Radial Frequency patterns), but not meeting these criteria in ShapeComp, did not necessarily register as perceptually uniform and circular. Therefore, leveraging ShapeComp, we introduce an automated method for generating extensive sets of perceptually uniform and circular shape spaces. We provide five newly validated circular shape sets derived from intricate naturalistic shapes, along with MATLAB code facilitating the creation of a limitless number of such sets. This advancement empowers cognitive scientists to construct large sets of perceptually uniform stimuli, allowing for a nuanced exploration of the impact of higher-level factors on object perception.