CUGIR

This dataset provides a single county subset of Microsoft's computer-generated building footprints. The original footprints, covering all 50 US States, were created via a two-stage process. In the first step, semantic segmentation, a deep neural network was used to recognize building pixels on Bing aerial imagery, which is a composite of multiple sources of varying dates. In the second step, building pixels blobs were converted into polygons. It is not perfect, particularly in dense urban areas. This dataset provides a relatively complete dataset of building outlines, especially useful for areas where official building footprints may not exist.