Revamping OldNYC: AI and OpenStreetMap Transformations

Over the past two years, I've quietly overhauled the OldNYC photo viewer, adding 10,000 historic photos and enhancing the site's accuracy and efficiency using modern AI tools and the OpenStreetMap ecosystem. OldNYC now boasts 49,000 photos, up from 39,000 in 2016, thanks to significant improvements in geolocation, OCR, and mapping technologies.

## Enhanced Geolocation

Previously, OldNYC relied on the Google Maps Geocoding API, which struggled with historical intersections that no longer exist. In 2024, I integrated OpenAI's GPT-4o to interpret complex location descriptions, allowing us to accurately place photos by extracting relevant intersections from text. This change enabled us to locate 6,000 additional photos, achieving a 96% accuracy rate in mapped images.

## AI-Powered OCR

The original OCR system, based on Ocropus, was effective but prone to errors. In 2024, I rebuilt it using GPT-4o-mini, which significantly improved text coverage and accuracy. The new system increased text coverage from 25,000 to 32,000 images and outperformed the old OCR in 75% of cases. By focusing on high-resolution images and avoiding context that could lead to hallucinations, GPT delivered superior results.

## Transition to OpenStreetMap

Due to changes in Google's pricing model, I switched from Google Maps to OpenStreetMap and MapLibre, which offered faster rendering, smoother zooming, and customizable map styling. This transition not only reduced costs but also allowed me to remove anachronisms like highways that didn't exist in the 1930s.

## Future Directions

Looking ahead, I plan to leverage AI to extract more information from images, such as identifying people or buildings. I'm also contributing to OpenHistoricalMap, which could simplify photo geolocation if it includes comprehensive historical street grids for NYC. Additionally, I'm interested in helping developers create similar sites for other cities. If you're interested, feel free to reach out.