Abstract
Curvature has been suggested to play a crucial role in supporting visual object processing and functional selectivity in high-level visual regions. While for artificial stimuli there may be a clear definition of curvature, this is more challenging for natural images, and the definition of curvature can vary from the curvature of the global shape of objects cropped from background to local elements of textures, both of which may deviate from our subjective percept of the curviness of individual natural images. How can we quantify perceived curvature of natural images, and how does this perceived curvature relate to patterns of brain activity? To improve our understanding of perceived curvature, we gathered extensive curvature ratings for 1,854 objects across 27,961 natural images of the THINGS database, compared their alignment with fMRI responses to computed curvature measures (Li & Bonner, 2020; Walther & Shen, 2014), and developed a neural network model that predicted perceived curvature for new images. Perceived curvature exhibited high reliability (r = 0.93). Computed curvature only weakly correlated with perceived curvature (r = 0.27 and r = 0.30) but also weakly correlated with each other (r = 0.22). In the human visual system, perceived curvature generally accounted for more variance across higher-level visual cortex than other measures and corresponded best to known category selectivities (e.g., Li & Bonner, 2020; Long et al., 2018). Given the validity of this curvature measure, we aimed at providing an automated quantification of perceived curvature for novel images. To this end, we finetuned a convolutional neural network to predict the perceived curvature of images, achieving notable performance (cross-validated R2= 64%). Together, our results highlight the importance of perceived curvature as a mid-level summary statistic and provide an approach for the automated quantification of perceived curvature in natural object images.