This is a cross-post from my blog
Microsoft released a machine-generated dataset of building footprints for the United States some years ago. The footprints are derived from aerial imagery. This works well, most of the time. Where you run into problems is in rural areas, especially where there’s natural features and topography that throws the machine learning off. Then the machine starts to think all kinds of things are buildings:

Contenedor de ropa usada en la avenida Las Caballerizas. Fuente: trabajo propio (