Built an app that guesses where you are in the world from a single photo using image embeddings. Started as a GeoCLIP-based prototype, evolved into a full pipeline — downloaded 51 million street view images and trained geospot-base, a 400M image-to-GPS embedding model.
Over 3 months I trained 449 different embedding models. geospot-base is trained on publicly available Flickr and Mapillary data, with geospot-pro achieving SOTA on the 750km benchmark.