Abstract
We know much more about neural representations of visual objects than we do about those of large extended surfaces in a scene, despite the importance of surfaces for navigation and evidence that object and scene representations are anatomically separate and may be subject to different attentional limits (e.g., Greene & Oliva, 2009). One particularly critical question about scene representation is the extent to which we represent specific configurations of major surfaces in natural scenes, beyond global properties like openness or naturalness. Recent work (Lescroart & Gallant, 2019) finds that information about the 3D configurations of major surfaces in scenes is detectable in scene-selective visual areas—however, this work was unable to tell whether this was driven by global, image-level summaries (e.g., whether an image as a whole has left-oriented surfaces) or more spatially localized scene surface information that could be useful in guiding our actions in the world (e.g., the right lower portion of an image has left-oriented surfaces). Indeed, this question is difficult to answer because of high degrees of covariation between these types of information across stimuli. In the present work, we use a stimulus selection procedure similar to Groen et al. (2017) to identify a stimulus set in which a global-3D scene surface model and a scene surface model capturing image-quadrant-level spatial specificity are minimally correlated (RDM correlation of -0.04). We assess the presence of each of these types of information using fMRI in scene-selective cortex by evaluating prediction performance on held-out voxel responses, and we find that both global- and quadrant-based- 3D scene surface information are present in about equal measure. Future work could examine the circumstances under which each persist in memory and/or during orthogonal tasks, with implications for representations that may help us navigate the world with our limited working memory capacity.