01Available
Geo-Estimation
Establishes where the photograph was taken.
Manta
Development previewManta
A visual evidence model for work that gets questioned.
Manta establishes where a photograph was taken, resolves it to the street or the building, shows the evidence that led there, and never claims more precision than that evidence carries.
Approved organizations · No individual tier
01 / The model
Precision,
bound to evidence.
Calling it a geolocation model undersells it: locating the photograph is where the work starts. Manta is for the cases where somebody will disagree with the answer.
- 01
Establishes
Where an image was taken, from its content alone, with no metadata.
- 02
Resolves
That answer down to the street or the building, rather than stopping at a radius.
- 03
Shows
The evidence that led there, so the conclusion can be argued with.
- 04
Holds to it
The precision it states limits what it may claim. No street in the evidence, no street in the answer.
02 / Three modes
One model.
Three jobs.
Each mode answers a different question and fails differently, with one evidence chain running through all three. You choose the one that fits the work.
02Available
Street-Match
Resolves the candidate regions to the exact street or building.
03Early research preview
Property-Verify
Confirms a photo matches a specific property, down to the room.
Why Property-Verify is last
It is the only one of the three that ends at a front door. We could switch it on in more places than we have. We have not, because a tool that names the building in a photograph unsettles people, and it should. It runs only inside a reviewed workflow, for organizations we have checked, with a person reading the result.
03 / Results
Measured, including
where it does not lead.
Street-Match is evaluated on seven place-recognition benchmarks; geo-estimation on the OSV-5M test split. The full tables, sources and protocol are in the announcement.
Street-Match · Nordland · Recall@1
97.21%
+1.21 over the best other reported model (SAGE (8448-D), 96.0).
Leads on 2 of 7 benchmarks. Ranked #2 of 11 by mean Recall@1. Numerical leads, not established improvements beyond evaluation variability.
| Benchmark | Street-Match | Best other | Margin |
|---|---|---|---|
| SPED | 95.39 | SAGE (8448-D) 98.9 | −3.51 |
| Pitts30k | 94.95 | SAGE (8448-D) 95.8 | −0.85 |
| MSLS val | 94.60 | SAGE (8448-D) 94.5 | +0.10 |
| Nordland | 97.21 | SAGE (8448-D) 96.0 | +1.21 |
| AmsterTime | 74.25 | SAGE (8448-D) 83.5 | −9.25 |
| Tokyo 24/7 | 98.43 | EffoVPR 98.7 | −0.27 |
| SVOX | 98.74 | BoQ 99.0 | −0.26 |
Geo-Estimation · OSV-5M · share within distance
Median error 30.5 km vs 32 km (Chipoint-2 card)
45.77%
Within 25 km
+0.94 vs Chipoint-2
81.83%
Within 200 km
+0.73 vs Chipoint-2
92.57%
Within 750 km
+0.46 vs Chipoint-2
97.30%
Within 2,500 km
+0.48 vs Chipoint-2
Numerical leads; the smaller differences should not be read as established improvements beyond evaluation variability.
04 / Access
Granted to an organization,
not a person.
Enterprise organizations first, then selected pilots and research partners, then wider access. Controlled access is a stage, not a destination. Oceanir’s standard model stays the default across the app and the API.
- 01
We verify the organization
It names the people who will use Manta and what they will use it for. The grant is scoped to that.
- 02
Coverage per city, per purpose
Enabled against a declared purpose, not switched on globally. Dense reference coverage today: Lisbon, San Francisco and Miami.
- 03
A reviewed workflow
Manta runs where a person reads the result and is accountable for what it concludes.
05 / Limits
Where it fails,
stated plainly.
- Interiors give it very little.
- Overcast skies remove the light it reads latitude from.
- A motorway in one temperate country can look a great deal like a motorway in another. Sometimes the honest answer is a region.
The people boundary
Manta answers where a photograph was taken. It is not built to find people, and we do not permit it to be used that way.
That is written into our terms and enforced in our test suite, not left to good intentions.
Send us your hardest cases.
Including the ones you already know the answer to. We return exactly what Manta produces, including where it declines to commit.