During State of the Map US Microsoft outlined it's target to leverage computer vision to create a building footprint indistinguishable from a human traced edit. We believe we have hit that milestone for at least 75% of the cases for the United States.

In this technical talk Microsoft will describe using Azure COSMOS DB in conjunction with the Open Source CNTK framework to develop feature extraction models and polygonization algorithms for footprint extraction at scale. We will present the relevance of building footprints as well as other potential targets for computer vision including coastlines, water features, surface=* tags and land use.