How GeoPin Works

Every image, both the query and the millions of reference images in our database, is processed by GeoPin's visual embedding engine, a specialised deep learning model for place recognition. It transforms each image into a compact 512-dimensional embedding vector that captures its geographic visual identity.

Unlike classification models that predict discrete locations, our model learns a continuous embedding space where visually similar places cluster closely together. Two photos of the same street corner, taken at different times and from different angles, produce vectors that lie close together in this space.

Vector similarity alone can produce false positives: different locations that happen to look alike. To eliminate these, GeoPin applies geometric verification using local feature matching.

GeoPin's verification engine extracts robust local features from both the query image and each candidate reference image. A specialised feature matcher then matches these features with a lightweight attention-based architecture, finding correspondences that satisfy the geometric constraints of a true scene match.

Only candidates with sufficient geometrically consistent feature matches pass verification. The final result includes GPS coordinates, a confidence score and the matched reference images.