Single Image Geolocation You Can Watch Work

Geolocating one frame with no metadata is a genuinely hard problem, and most public demos of it are curated. This one is not: it runs on whatever round the game deals, in front of you, and it is wrong about one time in five. If you work in computer vision, geospatial, or OSINT, that failure rate is the interesting part. Disclosure: ATLAS is our own product, and this page does not explain how it is built.

Rated 5.0 out of 5 from 11 verified buyers. Read every review.

Why a game is a good benchmark

GeoGuessr deals rounds nobody selected, from a distribution nobody tuned for, under a clock. That is a harder and more honest evaluation than a held out split of the same corpus the model trained on, and it is why the number here is about 81 percent rather than the high nineties you see quoted elsewhere in this space. Numbers in the high nineties usually mean a curated test set, or a tool that reads coordinates instead of predicting them.

There is also a scoring function attached, which is the part people miss. Getting the country right is not the same as getting the point right. The score is driven by distance, and around 4,000 points a round is the honest average.

The three failure classes

These are different problems and it is worth keeping them apart, because only one of them is solved by more data.

Feature poor scenes. An empty motorway, a stand of identical conifers, a field in fog. There is no signal in the frame. A human expert does no better here, and more training images do not create information that was never in the picture.

Visual convergence. Neighbouring countries that share road standards, sign conventions, vehicle fleets and vegetation look the same on an ordinary residential street. This is a discrimination problem inside a tight cluster, not a coverage problem, and it is where confident wrong answers come from.

Thin coverage. Street level imagery follows roads, camera cars and permission, so the dataset is a map of where a car was allowed to drive, not a map of the planet. Some regions are simply under photographed. This is the class where more images genuinely help.

What is worth watching for

• Confidence calibration. The interesting failures are the confident ones. Watch what it does on a residential street in a country cluster rather than on an obvious landmark.
• Latency under a real constraint. About 3 seconds is not an optimisation target chosen for a benchmark, it is the point past which the answer is useless in a timed round.
• Country versus distance. Two systems with the same country accuracy can differ by hundreds of points a round on distance. Country accuracy alone is a weak metric and it is the one everyone quotes.

Frequently asked questions

On GeoGuessr the desktop app predicts from the rendered frame. It does not read coordinates out of the page and does not touch the game internals, which is also why it can be wrong. The browser extension on the smaller clone games works differently: there it reads the round information the page already received, so no prediction is involved and no accuracy figure is claimed for those games.

About 81 percent correct country, across more than 1,600 real ranked rounds played in July 2026, over 111 supported countries. That is a live play figure rather than a curated test set, which is the reason it is lower than numbers you will see quoted elsewhere in this space. Scores average around 4,000 a round, and country accuracy is not the same thing as distance accuracy.

Three distinct failure classes, and they need different fixes. Feature poor scenes give the model nothing to work with, and no amount of extra data helps there. Visual convergence between neighbouring countries that share infrastructure standards is a discrimination problem, not a coverage problem. Thin coverage in under photographed regions is the one that more data would actually solve.

Yes. There is a public demo video, and the browser extension gives you 7 rounds free with no account. Watching it on rounds you did not choose is the only demo worth anything in this area, because a curated screenshot proves nothing about a geolocation model.

No. Neither the weights nor the training set are published or licensable, and this page does not describe the architecture. What is public is the behaviour: the accuracy figure and how it was measured, the latency, the country coverage, and the failure classes above.

More from ATLAS

How ATLAS predicts a location is the conceptual explainer, and how accurate are GeoGuessr AI tools covers the measurement method and why the high claims in this category do not hold up. The eight platform repositories on GitHub are the public surface for each build.