BinoML: Supervised ranking for automatic building labeling
2022
Building numbers shown on building outlines of a map are important information for guiding delivery associates to the correct building of a package’s recipient. Intuitively, the more labeled buildings are present in our map, the less likely to misplace an order in addition to other benefits such as delivery efficiency as drivers get better visual cues about building positions. Although there are free and collaborative projects for creating geographic database of the world, such as the OpenStreetMap (OSM) [2] which also supplies building outlines along with their building numbers, many building outlines still remain unlabeled in many U.S. regions and other countries. Hence, we are interested in developing models that can automatically add building numbers with ≥ 99% precision to unlabeled buildings across geographies with low to medium building number coverage. In this paper, we describe a ML model which in offline results showed 2% to 12% increase in building number coverage in some US regions compared to that of the OSM. The proposed model can also be applied to improve the building number coverage of other countries after fine-tuning to those new regions.
Research areas