Using OpenStreetMap to create training data
summary
OpenStreetMap features become labeled satellite training data by generating an Overpass query from a config, downloading tiles and exporting annotations as MS COCO JSON using ohsome2label, then cleaning and editing with CVAT and exploring/filtering large datasets with FiftyOne. Common surprises include missing or partial imagery, cloud cover, geolocation shifts between map and imagery, and incorrect or incomplete OSM tags; workflows include manual review, adding missing labels, and hard negative mining to improve model precision and guide where mapping updates are needed.
AI-written summary · about
related talks
01h 41mOpenStreetWhat? Mapping The World With Open Data
22mIsland Mapping: Measuring Urban Footprints with Mapillary & Open Source Tools
15mBeyondTracks: an OpenStreetMap case study
22mUnderstanding our cities with FOSS4G
13mUsing open source mapping tools to understand indoor navigation routes...
31mBuilding a Scalable, Reproducible Ontology System to Improve Deep Learning Data
19mHow we leverage AI to build great maps at massive scale
48mEmpowering AI Innovation Through Open Geospatial Ecosystems