Where Was This Photo Taken? AI Instantly Identifies Location
AI Can Now Geolocation Photos with High Accuracy, Speed, and Efficiency
This article details a new AI model developed by researchers at China University of Petroleum that accurately geolocates images by matching street-level photos to aerial views. Here’s a breakdown of the key takeaways:
Key Features & Performance:
* High Accuracy: Achieves up to 97% accuracy in the initial location narrowing stage (with 180-degree field of view images), comparable to or better than existing models. Pinpoints exact locations with 82% accuracy, within three percentage points of competitors.
* Speed & Efficiency: At least twice as fast as similar models and uses less than a third of the memory. This makes it suitable for applications like navigation systems and defense.
* Deep Cross-View Hashing: The model utilizes a technique called deep cross-view hashing, which transforms images into unique numerical “fingerprints” for faster comparison.
* Vision Transformer: Employs a deep learning model (a vision transformer – similar architecture to ChatGPT, but for images) to identify key landmarks and encode them into thes number strings.
* Averaging Technique: After identifying potential aerial matches, the system averages the geographical data, weighting closer locations more heavily to minimize errors.
How it Works:
- Image Encoding: The AI analyzes street-level photos, identifying key landmarks (buildings, fountains, roundabouts) using a vision transformer.
- Hashing: These landmarks are converted into unique numerical codes (hashes).
- Matching: The code from the street-level photo is compared to the codes of aerial images in a database.
- Location Estimation: the five closest aerial matches are identified, their geographical data is averaged, and an estimated location is determined.
Expert Opinions:
* Peng Ren (China University of Petroleum): The AI is trained to focus on key landmarks, creating a shared “language” between street and aerial views.
* Hongdong Li (Australian National University): The numerical code acts like a fingerprint, capturing unique features for swift geolocation.
* Nathan Jacobs (Washington University in St. Louis): While a clear advance, the problem has been solved before and he doesn’t consider it groundbreaking.
Publication:
The research was published in IEEE Transactions on Geoscience and Remote Sensing.
In essence, this new AI model offers a notable improvement in geolocation technology due to its combination of accuracy, speed, and memory efficiency.