Offline geolocation from a map screenshot
OCR, a gazetteer and a road graph instead of cloud services
Locates a map screenshot within a hard 15-second budget, fully offline. When the evidence is thin, it honestly answers with a region or "unresolved" instead of a confident mistake.
One real run
From map to place
On one predeclared, real, openly licensed MAP_2D sample, the offline pipeline returned “probable Amsterdam.”
Input · MAP_2D

Map of central Amsterdam, selected before the pipeline run.
Amsterdam centre map.png by Dr. Blofeld / Wvdp; © OpenStreetMap contributors; CC BY-SA 2.0; via Wikimedia Commons.
Map pixels are preserved; responsive WebP copies are generated for display. The copies use the same license.
Static projection of the result
Land outlines: Natural Earth via world-atlas 2.0.2. Natural Earth terms · world-atlas 2.0.2 (ISC).
The marker shows the pipeline output, not measured point accuracy.
Pipeline result
Probable Amsterdam
Netherlands · North Holland
- This offline run
- 2.34 s / 15 s
- OCR
- RapidOCR
- Network during the run
- Blocked
Among the labels recognized
- Amsterdam
- Amsterdam-Centraal
- Centraal
- Oosterdok
- Kattenburg
- Grachtengordel
- Oosterpark
- Museumplein
Verifiable decisions
Map labels alone never produce an exact point. The answer is a region, a street with a radius, or "unresolved".
code
In a local run on 2026-09-18, p95 was about 750 ms against a 15-second budget.
measured run
In detail
The job
Given a map screenshot, find the place within a hard 15-second budget, with no network and no cloud services.
How it works
ONNX-based OCR reads the labels, a world gazetteer finds candidates, and country hints are combined. Each candidate is then checked against the OpenStreetMap road graph: road names and numbers near the point, plus junction topology taken from the screenshot. Indexes live in SQLite with full-text search and R-trees, so a graph query takes milliseconds.
Honest uncertainty
Text alone never yields an exact point. The answer is a region, a street with a radius, or “unresolved”. An honest region beats a confident mistake.
What is measured and what is not
Latency in a local run on 2026-09-18: p50 about 560 ms, p95 about 750 ms against a 15-second budget. On one predeclared, real, openly licensed MAP_2D sample, the offline pipeline returned “probable Amsterdam.” This demonstrates one result; it does not measure accuracy. An accuracy benchmark on a set of real labelled screenshots is not published.
Data: GeoNames (CC BY 4.0) and OpenStreetMap (ODbL).