Search all text in New York City
A new project indexes text from millions of New York City Google Street View panoramas, letting people search everything from graffiti tags and storefront signs to political slogans and odd OCR errors. Commenters are delighted by the playful searches and cultural Easter eggs it enables, while also noting the limitations of the text recognition and the heavy reliance on Google’s Street View data and pricing. Many see potential extensions, from vector-based visual search and OSINT-style analysis to similar tools for other cities and use cases like urban research and infrastructure monitoring.
Overall reception and uses
- Many commenters find the project delightful and “exceedingly fun,” describing it as something they could spend hours exploring.
- People immediately use it to find personal landmarks (e.g., childhood bagel shops) and local culture (graffiti writers, stickers, slogans, political posters).
- Some note its value for OSINT and imagine that intelligence agencies likely have similar tools at global scale.
Playing with the search
- Users test funny or crude words (“fart,” “pedo,” “sex,” “foo,” “fool”), getting amusing misreads and coining it as a kind of game.
- Another game emerges: find real English words with the fewest hits; examples like “scintillating,” “calisthenics,” “perplexed,” “Buxom,” etc.
- People search for graffiti tags, politicians’ names, slogans, and niche phrases to probe cultural traces across the city.
- Food terms (“bagels,” “pizza,” “sushi,” “hotdog,” “massage”) reveal dense and uneven spatial distributions; one person notes sushi is heavily Manhattan‑centric.
OCR quality and quirks
- Multiple comments say the idea is brilliant but current OCR accuracy is “pretty bad” for many queries.
- Misreads of Google watermarks, cropped signs, and partial words generate large numbers of false positives.
- Some searches work well for clear signage; others show systematic errors: “OPEN” → “OBEY,” “food” → “foo,” and numerous comical reinterpretations.
Technical and cost considerations
- Commenters estimate OCR compute as manageable on consumer hardware, but highlight Google Maps / Street View API costs (tens of thousands of dollars at list prices) as the real barrier.
- Discussion notes ~8 million panoramas processed; various back‑of‑the‑envelope calculations of image throughput and API fees appear.
- A linked talk suggests the creator used publicly-available Street View imagery and macOS’s built-in OCR via Shortcuts, possibly without paid API access; it’s unclear how rate limits were handled.
Related projects and desired extensions
- Links to similar efforts: earlier Brooklyn‑only and London versions, a New York traffic‑camera semantic search project, and a UK building‑safety use of Street View.
- Several people want an API, deduplication of near-identical views, CLIP/semantic image embeddings, or a “text‑only Street View.”
- Others imagine this as a Google Maps layer for discovering niche businesses by sign text.
Data freshness, filtering, and tangents
- Some try to infer the capture timeframe from protest posters and political signs.
- There’s curiosity about why some official notices or offensive words are hard to find and speculation around mild censoring in the write‑up’s links.
- One tangent raises the lack of simple, accessible text‑to‑speech tools for blind users; replies point to cost and existing assistive tech rather than this project specifically.