How we check a photo location result
A location estimate is a claim to check. Our examples separate the input, the model's response and what a reviewer could establish from independent evidence.
Written and reviewed by the Oceanir team. Updated 2026-10-05 (UTC).
The recorded London scan
We submitted the existing London sample to the same public Surface endpoint used by the upload tool. The upload filename was sample.jpg. The request included D1 depth and worldwide search scope; it included no city caption, coordinates or location hint. We retained the response and input fingerprints.
- Recorded at
- 2026-10-05T02:24:06.699669+00:00
- Reported model
- Orca 1.5-preview. This is the API's returned label; its underlying provider version was not exposed.
- Image SHA-256
- 44693a833b907066a538f42eb2aa10bc8905fcb6b199658bc63b2c83c211c97a
- Response SHA-256
- cd36e3d4d558932bcfeb2e52541a46326f5cf2390bd03e360c244c69a84e5fd1
What counts as a verified match?
The Gherkin and London buses support a London hypothesis. The returned city agrees with that reference. Establishing the camera position requires a matching street layout, neighboring facades and viewpoint. We have not established those for this scan, so we report city agreement and leave the exact viewpoint unresolved.
Whittington Avenue was named by the model. It is an unverified suggestion. The returned confidence score does not independently validate it. A map marker, available Street View coverage or a matching country does not by itself establish the photographed street.
For a distance-error claim, we require an independently known camera coordinate. A landmark coordinate or a city center is insufficient. This example has no measured camera-distance error.
Selection and limits
This is one existing landmark photo selected to demonstrate the workflow. It is not a random test set or an accuracy benchmark. Results may change between runs and model versions. We cannot infer ordinary-street, interior or worldwide accuracy from this example.
The Chicago walkthrough remains an illustration of visual clues. Its follow-up scan was rate-limited and produced no model result. A rate limit is a service outcome, not an inaccurate location prediction. The public scan has usage limits; a visitor can inspect the recorded London example before deciding whether to continue.
A future evaluation should record the model and mode, image provenance, independently established reference positions, sample selection and size, all predictions and failures, and distance thresholds. Any fitting or tuning must be separated from the reported test images.
Before using paid review
Try the free scan on an image with visible surroundings. Check its proposed place against independent imagery. Paid review adds analysis and review tools according to your plan; it can still return a broad area or an unresolved result. Evaluate whether those tools help with your image before subscribing.