Low-level visual semantic (left): structure constraints between neighboring pixels; middle-level visual semantic (middle): structural regions corresponding to façades, footprints, and roofs; ...