Rainbolt is one of the world’s best players of Geoguessr – a game in which you are given a 360 picture and you get about 20 seconds to guess where in the world it was taken. A team at Stanford took 2 months and built an AI that can guess 92% of countries correctly and a median miss error of only 44km – which is astounding.

Here’s a head-to-head competition between the AI Predicting Image Geolocations (or PIGEON) and a pro geogussr player:

But there’s another side of this kind of technology. NPR did an interview and presented a few personal photos to the algorithm. PIGEON was able to guess the location the photo was taken to a really high degree of accuracy. This means you can find places taken in old family snapshots, but it also means that algorithms like this can reveal everywhere you are, and have been, based on your social media posts.

How it works

The algorithm that PIGEON uses is an interesting combination of AI model techniques. Besides the AI based learning, they use some interesting methods such as ‘geocells’ that uses political/geographic regions to help narrow locations instead of just naïve squares.

Rainbolt even pointed out that PIGEON picked up on camera lens smudges in the sky that were very common in Canadian google image captures:

There’s so many other details. Definitely check out their paper here: https://arxiv.org/abs/2307.05845

Summary

This is yet another example of where 3 graduate students were able to develop a system that is better than the best experts in the world. And in this case, they did it in less than 3 months with off the shelf software and hardware.

You can only imagine where things will be in just a few years. Anyone that doesn’t think AI is already changing the world is missing it as it’s happening.