One photograph, one country
This page uses a different model from the ten-view analyser. That one is trained on synthetic panoramas of invented countries and demonstrates the multi-view pipeline. This one is trained on 403,807 real street-level photographs from OpenStreetView-5M and predicts among 71 real countries.
Classify a street photograph
One photograph, one country. The image is resized to 224 px and stripped of all metadata on the server before the model sees it — EXIF, GPS and camera details are discarded, never stored and never sent to the model.
Drop a photograph here
JPEG, PNG, WebP or AVIF · up to 12 MB
Countries this model knows71
Read from the checkpoint itself, so it cannot drift from what the model actually predicts. A photograph from anywhere else is forced into the nearest of these.
- Albania
- Argentina
- Australia
- Bolivia
- Brazil
- Canada
- Cape Verde
- Chile
- China
- Colombia
- Congo - Kinshasa
- Croatia
- Egypt
- El Salvador
- Ethiopia
- Fiji
- Finland
- France
- Germany
- Ghana
- Greece
- Guatemala
- Hungary
- Iceland
- India
- Indonesia
- Iran
- Iraq
- Ireland
- Italy
- Japan
- Kazakhstan
- Madagascar
- Mali
- Mexico
- Micronesia
- Morocco
- Mozambique
- Nepal
- Netherlands
- New Zealand
- Nigeria
- Pakistan
- Papua New Guinea
- Peru
- Philippines
- Poland
- Portugal
- Russia
- Saudi Arabia
- Senegal
- Serbia
- Sierra Leone
- Slovakia
- South Africa
- South Korea
- Spain
- Sweden
- Tajikistan
- Tanzania
- Thailand
- Togo
- Tunisia
- Türkiye
- Turkmenistan
- Uganda
- United Kingdom
- United States
- Uzbekistan
- Vanuatu
- Zambia
What this model can and cannot do
- Test accuracy
- 58.6% top-1 across 71 countries
- Macro-F1
- 0.4497
- Calibration error
- 0.0790
- Encoder
- convnext_tiny, 224 px
It is right rather more than half the time. Chance across 71 countries is 1.4%, so 58.6% is roughly forty times chance — and still wrong two times in five. The confidence figure is calibrated, which means it is worth reading: when the model says 70% it is right about 70% of the time.
It is uneven. Macro-F1 of 0.45 against 58.6% top-1 means a minority of countries carry most of the accuracy. Coverage in the training data is heavily skewed toward where contributors drive, so a country represented by a few hundred images in one city is barely learned at all.
It only knows 71 countries. A photograph from anywhere else will be forced into the nearest of those, confidently or otherwise. Countries occupying too small an area to split into separate training and test regions were excluded entirely, which removes most city-states and small island nations.
It works on street scenes. It learned from road-level photographs — signage, markings, poles, kerbs, vegetation, architecture. Interiors, close-ups and aerial shots are outside what it has seen, and it has no way to tell you so beyond a low confidence.
A prediction is a statement about what a place looks like. It is not evidence of where a photograph was taken, and it is certainly not evidence about any person. Training imagery is CC BY-SA 4.0 from OpenStreetView-5M contributors. Full details in the model card.