Abstract
A network of areas in the human brain—including the Parahippocampal Place Area (PPA), the Occipital Place Area (OPA), and the Retrosplenial Cortex (RSC)—represent visual scenes. However, it is still unclear whether these areas represent high-level features (such as local scene structure or scene category), or low-level features (such as high spatial frequencies and rectilinear corners). To better characterize the representation of visual features in PPA, OPA, and RSC, we define several feature spaces hypothesized to be represented in scene-selective areas, and use them to predict brain activity. The feature spaces we define reflect scene structure (3D orientation of large surfaces), scene expanse (distance from visible surfaces to the virtual camera), and local spatial frequency (Gabor wavelet transformations) of each rendered scene. We fit each feature space to BOLD fMRI data recorded while human subjects viewed movies of a virtual world rendered using 3D graphics software. The graphics software allows us to quantify scene structure and expanse with continuous parameters instead of categorical labels such as “open” or “closed”. We fit models for all feature spaces using L2-regularized regression, and evaluate each model based on how much response variance it predicts in a withheld data set. The scene structure model explains more response variance in our data than the scene expanse and local spatial frequency models. A hybrid model that combines surface orientation and distance from the virtual camera explains still more variance. These results are consistent whether subjects attend to the size of each scene or to the number of objects in each scene. Together, these results suggest that PPA, OPA, and RSC represent conjunctions of the depth and orientation of walls, ceilings, and other large objects in scenes.
Meeting abstract presented at VSS 2015