In a groundbreaking initiative called Predicting Image Geolocations (PIGEON), three Stanford graduate students, Michal Skreta, Silas Alberti, and Lukas Haas, harnessed the power of artificial intelligence to excel in geolocating photos, demonstrating its potential and raising privacy concerns.
The project originated from a Stanford class, Computer Science 330, where the students, avid players of the online game GeoGuessr, sought to build an AI player surpassing human capabilities. Using the OpenAI-developed CLIP neural network, they trained their system with a dataset of 500,000 Google Street View images, achieving remarkable performance. PIGEON can correctly identify the country in 95% of cases and pinpoint locations within approximately 25 miles.
Silas Alberti emphasized the program’s ability to surpass human geoguessers, recounting their victory against renowned human geoguesser Trevor Rainbolt. PIGEON’s success is attributed to its capacity to discern subtle visual cues, including variations in foliage, soil, and weather, making it adept at image classification by global position.
The potential applications of PIGEON are extensive, ranging from identifying infrastructure issues to aiding biodiversity monitoring and serving as an educational tool. However, the project raises concerns about privacy, voiced by Jay Stanley from the American Civil Liberties Union. Stanley warns of potential misuse for government surveillance, corporate tracking, or even stalking, emphasizing the sensitivity of location information from a privacy standpoint.
To test PIGEON’s capabilities, the program was challenged with personal, unpublished photos from a cross-country trip. Impressively, it accurately identified a campsite in Yellowstone within 35 miles and a street in San Francisco within a few city blocks. While some mismatches occurred, the ACLU’s Jay Stanley acknowledges the project’s overall potential, pondering what larger entities, like Google, could achieve.
The conversation extends to existing features, such as Google’s “location estimation,” utilizing AI to guess photo locations. While currently limited to a catalog of approximately a million landmarks, concerns arise regarding potential expansion for tracking user travel or government surveillance. Stanley emphasizes the shift in removing GPS location tags from photos may become ineffective.
Despite their achievement, the Stanford students recognize the risks associated with AI geolocation and have refrained from publicly releasing their full model. Stanley anticipates the growing influence of AI in geolocation but emphasizes individual awareness of the content in background photos posted online.
The PIGEON project, born from a student endeavor, showcases the remarkable capabilities of AI in geolocating images, opening doors to diverse applications. However, the ethical implications and potential for privacy invasion underscore the need for responsible development and usage of such technology.