Reasons for sub-optimal performance of visual CNN models