Representation of object contour features in intermediate visual areas in the human brain
53.427, Tuesday, May 14, 8:30 am - 12:30 pm, Orchid Ballroom
Mark D. Lescroart1, Shinji Nishimoto1, Jack L. Gallant1; 1University of California, Berkeley
Vision is mediated by a set of hierarchically organized cortical areas that represent information at different levels of complexity. Peripheral visual areas represent simple image-level features such as oriented edges, textures and motion energy. Central visual areas represent the semantic categories of objects and scenes. Intermediate areas are thought to represent shape features (such as silhouette contours) that are critical for building up more complex representations from simpler features. However, because it is difficult to parametrize shape features in natural photographs, most studies of shape representation use simple objects on blank backgrounds. Consequently, intermediate representations of complex shapes in natural scenes are still poorly understood. To investigate this issue we used computer animation software (Blender) to render realistic artificial scenes that varied in color, texture, and lighting, as well as in object position, size, and semantic category. Human subjects viewed the rendered movies for 70 minutes while BOLD fMRI responses were recorded from visual cortex. We then computed contour features associated with objects (i.e., silhouettes and surface/depth discontinuities within objects) based on the virtual 3D world that generated each scene. We used these features as regressors in a voxel-wise modeling and decoding (VWMD) framework. VWMD estimates a set of weights for each individual voxel that reflect how specific features influence hemodynamic responses. Finally, we used a separate data set to test predictions of the fit models. We found that predictions of the object contour models were more accurate in V4 and lateral occipital cortex than were predictions of a simple motion energy model shown previously to provide good predictions in V1. Furthermore, predictions did not improve when additional regressors that reflected contours at surface/depth discontinuities in the background were included. Our results suggest that intermediate visual areas represent the contours of foreground objects in the natural world.