Seohee Han¹ (), Dirk B. Walther²; ¹University of Toronto

Detecting and integrating contours that delineate the boundaries of objects, surfaces, and other scene elements is a crucial function of what is loosely called “mid-level vision.” How are such contours and their properties processed by the brain? We here explore this question using the high-resolution Natural Scenes Dataset (Allen et al, 2021). We analyzed the BOLD activity related to eight participants viewing subsets of 73,000 images of objects and scenes within V1, V2, V3, and hV4. Using the population receptive fields of individual voxels, we sample contour properties in a spatially specific manner to construct individual regressors for each voxel. This technique, first described by Roth et al. (2022), allows us to determine to what extent voxels within the visual cortex contribute to the representation of contour properties within their receptive field across thousands of images. When analyzing the salient contours in the images, we find a strong preference for horizontal orientations, consistent with the importance for scene layout, such as the horizon line. Interestingly, this finding contrasts with a similar analysis that relies on analyzing the orientation-specific Fourier energy in the photographs, which showed a primarily radial organization of orientation preference across the visual field. We present direct comparisons of the two methods. The technique of sampling contour properties with spatial specificity opens the door to exploring a range of other contour properties, such as contour curvature, contour junctions, as well as relationships between contours, such as parallelism and symmetry. Observing the neural representations of these properties and relationships across visual regions will bring us closer to a mechanistic understanding of how our perceptual information is organized in mid-level vision.