Winterist/rs-geolocation-benchmark

Remote-sensing geolocation benchmark

Per-image view of a remote-sensing geolocation benchmark: pick one of 1000 Sentinel-2 images and see where 18 models thought it was, how far off each one was, and what each of them said while reasoning.

• —2D / Globe — the flat map keeps every prediction on screen at once; the globe (CesiumJS) reads better when errors span continents. Both draw the same Esri satellite imagery and the same great-circle arcs, so either can be zoomed in to inspect the ground.
• —The map is centred on the ground truth; the dashed circle is the area the image actually covers (109.8 km tile, 77.64 km half-diagonal).
• —Marker colour encodes the error band. Hover a marker to see which model it is, click it to jump to that row.
• —The table lists every model's predicted coordinates, error and full reasoning, rendered as Markdown.
• —Click the satellite image to view it full size.

This is a static snapshot exported from the evaluation results — no inference runs here.

Imagery: Copernicus Sentinel-2 (L2A), processed for display. Basemaps and 3D imagery: Esri, Maxar, Earthstar Geographics.