Towards Total Scene Understanding: Classification, Annotation and Segmentation in an Automatic Framework