Microsoft released a comprehensive dataset for computer-generated building footprints in the United States. The method:
We developed a method that approximates the prediction pixels into polygons making decisions based on the whole prediction feature space. This is very different from standard approaches, e.g. Douglas-Pecker algorithm, which are greedy in nature. The method tries to impose some of a priory building properties, which are, at the moment, manually defined and automatically tuned.
The GeoJSON files for each state are available for download, released under the Open Data Commons Open Database License. Nice.