Street-View Geolocation Transformer (S.V.G.T.)
S.V.G.T. estimates latitude and longitude from street-view images. The visual transformer encodes each perspective view, ranks geographic cells, and predicts a coordinate within each candidate cell. A single fixed-view screenshot can be used as input.
Interactive demo
Gradio demo supports Chinese and English, 1–4 views, and ranked coordinates with Google Maps links. Run it locally with a separate checkpoint or deploy the demo folder as a Gradio Space.
Local inference
Use Python 3.10 or later and a compatible local S.V.G.T. checkpoint obtained separately. This repository contains the inference source and configurations.
python -m pip install -r requirements.txt
python -m streetgeo.inference.predict street_view.jpg --checkpoint /path/to/model.pt --device auto --top-k 5 --json
Run the command from this repository's root. It prints candidate coordinates, cell probabilities, and map links. For the optional country and region coverage filter, add --google-coverage-only. The bundled polygons provide a country and region approximation of coverage. See map source metadata.
The checkpoint carries the architecture configuration. The 400M configuration and legacy 135M configuration are included for reference.
The source snapshot comes from GitHub commit 3410436. The full project contains the training workbench, Windows client, and tests.