Abstract
Segmenting objects from backgrounds is a critical function of the human visual system. Studies have shown that intermediate-level visual areas including Lateral Occipital cortex (LO) represent information about object boundaries, but also represent other other features such as motion and visual semantic categories. However, because most studies use tightly controlled stimuli (e.g. stimuli with contours and no motion or vice versa), it is unknown how important object boundary contours are relative to other features. To address this issue, we measured fMRI responses while humans subjects viewed natural movies. We modeled responses to these movies as a function of motion energy, object boundary contours, and visual semantic categories. We used a customized version of a recent convolutional neural network (Deeplab v2) to label contours. We used the encoding models for each type of feature to predict withheld fMRI data, and used variance partitioning to determine whether the models explained shared or unique variance. We found that the boundary contour model explained some variance that was not explained by the motion energy model in LO. The same was true in areas selective for faces and bodies. However, this variance was shared with the semantic category model, and the motion energy model explained more unique variance than the boundary contour model. Thus, LO and other areas may represent boundary contours (perhaps specifically boundaries of animate entities), but that representation is mixed with representation of motion and semantic categories. Other features appear to affect responses in LO as much as object boundaries do.