What 5.3 Million Street View Images Taught Us About the World

The interesting thing about a 5.3 million image street level dataset is not its size. Size is easy to quote and tells you almost nothing. The interesting thing is that the world is not evenly represented inside it, and that the unevenness is not random. It follows roads, camera cars and permission, and it shows up directly in how well any system built on it performs.

This is what we learned looking at that data, framed as observations about the world rather than about the model. If you want the model side, how ATLAS predicts a location covers it.

Coverage is not the same as the world

Street level imagery exists where a vehicle with a camera on the roof was able to drive and allowed to publish. That is a much narrower set of places than it sounds. It excludes anywhere without a drivable road, anywhere a mapping vehicle was not permitted, anywhere too remote to be worth the trip, and large parts of several countries for reasons that are political rather than practical.

So the dataset is a map of vehicle access and permission. Reading it as a map of the planet is the first mistake available, and it is the one that makes people expect uniform performance. Places with no coverage do not exist to a system trained this way. Not underrepresented, absent. The 111 supported countries figure is a coverage fact before it is a model fact.

The long tail is most of the map

A small number of heavily photographed countries account for a large share of the images, and the tail behind them is long and very thin. This is the single most consequential property of the data and it has two consequences that pull in opposite directions.

The first is that a system trained naively on it becomes confidently biased toward the well photographed places. Shown something ambiguous, it reaches for the country it has seen most, and it is right often enough that the average score looks fine while the behaviour is wrong.

The second is that a benchmark drawn from the same distribution flatters you, because the test set is dominated by the same countries the training set is. You can improve a number like that for weeks without the tool getting better at the rounds people actually lose. Correcting for the imbalance makes the headline number go down and the product go up, which is an uncomfortable trade to make and the right one.

What makes a place easy

Two conditions have to hold at once, and people usually only think of the first.

It has to be visually distinctive. Something in an ordinary frame has to be characteristic: the script on a sign, the colour and pattern of road markings, the shape of bollards and guard rails, the side of the road traffic drives on, the species mix in the vegetation, the style of utility poles. These are the same cues a strong human player learns, and the tips post catalogues them by hand.

And it has to be well covered. A country can be extremely distinctive and still be hard, simply because there are not enough images of it. Distinctiveness without coverage is a place you would recognise if you had ever seen it.

Where both hold, performance is very good. Where either fails, it is not, and the two failures look identical from outside while needing completely different fixes.

What makes a place hard, and why

Three separate causes. Keeping them apart matters, because only one of the three is a data problem.

Thin coverage. Not enough images of a region for anything to be learned reliably. This is the one where more data genuinely helps, and it is the reason coverage gaps translate so directly into accuracy gaps. It is also the most fixable and the most boring.

Visual convergence. Neighbouring countries that share infrastructure standards, sign conventions, vehicle fleets and climate look the same on an ordinary residential street. There is nothing in the frame that separates them, so this is not a coverage problem at all. Adding a million more images of two countries that genuinely look alike does not make them look different. This is where confident wrong answers come from, and it needs something other than more data.

Feature poor scenes. An empty motorway with a treeline. A stand of identical conifers. A field in fog. There is no signal in the picture. A human expert does no better here, and this is worth saying plainly because it is often mistaken for a model weakness: the information was never in the frame. No amount of data creates it.

The numbers this produces

Put together, that data gives about 81 percent correct country across 111 countries, measured over more than 1,600 real ranked rounds in July 2026, in about 3 seconds a round, averaging around 4,000 points. The best single round on record is Tokyo, 11 km off, 4,987 out of 5,000, which is the ceiling rather than the average and should be read that way.

One distinction gets lost constantly and it is the one that matters most. Getting the country right is not the same as getting the point right. The score is driven by distance, so two systems with identical country accuracy can differ by hundreds of points a round. Country accuracy is the metric everyone quotes because it is the easiest one to state, and on its own it is close to useless for comparing tools.

What we would want more of

An honest wish list, framed as data rather than roadmap.

• Denser coverage in the thin regions. The most valuable images are of the places with the fewest, and they are the hardest ones to get, which is why the gap persists.
• More seasonal and time of day variety. A region photographed once, in summer, at midday, is a region you only recognise in summer at midday. Bare trees and low winter light change a scene more than people expect.
• Deduplication by location rather than by file. Near duplicates a few metres apart are different files and the same place. If they land on both sides of a train and test split, the benchmark quietly lies to you and you find out months later.

None of that closes the convergence problem, and it is worth conceding the limit. Two countries that share a road standard and a climate will keep being confused for each other, by software and by people, and the honest way to handle it is to say so rather than to promise a number that assumes it away. What got thrown away along the way covers the decisions that came out of accepting that.