A group of three graduate students at Stanford built an AI model that is capable of identifying the location of photographs by simply looking at the picture.
The developers of the AI claim the model, which has been trained on Google Street View data, is capable of correctly predicting the location of where a photograph was taken with 95% accuracy and within 25 miles of the real location. The team noted that this new AI model has real-world applications rather than just being used to track the location of people, as the new model could be used to promptly identify roads with downed power lines or biological surveys of regions.
The trio of engineers got the idea for the new AI model by being fans of the online game GeoGuessr, which places a player somewhere around the globe using Google Street View. The player is then required to guess their location by dropping a pin on the map, and the closer the players get to their actual location, the higher the score they get. To create the new AI model officially called "Predicting Image Geolocations" (PIGEON), the team of engineers used OpenAI's CLIP neural network, which learns about images through text.
"We created our own dataset of around 500,000 street view images," Silas Alberti, one of the Stanford students who developed the tool, told NPR. "That's actually not that much data, [and] we were able to get quite spectacular performance."
"From a privacy point of view, your location can be a very sensitive set of information," Jay Stanley at the American Civil Liberties Union told NPR.