Geolocalization Stockholm

Predicting where a street-view photo was taken in Stockholm using only its pixels, by comparing it against a gallery of images with CLIP, a general-purpose off-the-shelf embedding model. Toggle between a smaller and larger image gallery. The gallery is drawn from a dataset of 100K images I collected in preparation for my master's thesis.

Click the button (or reload the page) for a random query image, shown leftmost, alongside its 5 nearest neighbors from the gallery. On the map, blue dots mark those matches and the red circle marks the query's true location. Have fun!

Source and write-up: Stockholm-Geolocalization-Retrieval on GitHub. Also viewable directly on Hugging Face Spaces.