15 pixels square over every roof section nook of a raster image blended with the output of the 2D segmentation, and the category inference consist..