The OldNYC historic photo map grew from 39,000 to 49,000 photos through three major improvements made in 2024. GPT-4o was used to extract location data from photo descriptions, locating ~6,000 additional photos and raising the geocoding accuracy to 87%. OCR was rebuilt using gpt-4o-mini, dramatically improving transcription quality over the old Ocropus pipeline. The site also migrated from Google Maps to OpenStreetMap vector tiles and MapLibre to avoid ~$35/month in new Google Maps fees, while gaining the ability to customize the map to remove modern infrastructure like highways that didn't exist in the 1930s.

4m read timeFrom danvk.org
Post cover image
Table of contents
Better Geolocation with OpenAI and OpenStreetMapAI-Powered OCRMoving from Google Maps to OpenStreetMapWhat’s Next

Sort: