Geo-estimation
Where a photograph was taken, from the image alone. No GPS, no metadata, just the pixels. One of Manta’s three arms.
Process
How it works.
Four steps from photograph to answer. Each one narrows the candidates, and the last one decides how precise the answer is allowed to be.
Upload
Any photograph. No metadata required, and none is used.
Read
Architecture, planting, road markings, signage, light.
Reason
Weighs every cue against the others and against the candidates.
Answer
Ranked candidates, the evidence for each, and the precision it can defend.
Benchmarks
Accuracy by distance.
OSV-5M, the standard open-world geolocation benchmark: the share of photographs placed within each distance of where they were taken. Higher is better. Compared with Chipoint-2’s self-reported card, the strongest published result.
Within 25 km
Chipoint-2: 44.83%
Within 200 km
Chipoint-2: 81.10%
Median error
Chipoint-2: 32 km
The classical splits.
The older benchmarks the field still reports, scored at five distances. These are our rows; the full field, with every method we compared against and where we trail, is in the Manta post.
| Benchmark | 1 km | 25 km | 200 km | 750 km | 2500 km |
|---|---|---|---|---|---|
| Im2GPS3k | 22.52 | 50.82 | 66.57 | 80.91 | 91.16 |
| YFCC4k | 35.04 | 47.39 | 58.50 | 73.15 | 84.56 |
| GWS15k | 2.39 | 15.23 | 45.67 | 78.00 | 93.62 |
Percent of photographs within each distance of the true location.
Signals
Every photo is a clue.
What it reads. No single cue decides an answer, and how much each one counts depends on the photograph: a street sign can settle a city on its own, while a field of wheat says little more than a latitude.
Building styles, materials, roof shapes.
Lane markings, kerb profiles, signage conventions.
Tree species, canopy structure, ground cover.
Makes, plates, which side of the road they drive on.
Language, script, typeface, colour and placement rules.
Sun angle, shadow direction, haze.
Glossary
The terms, defined.
Visual location verification is a specific type of data reasoning. It removes the dependency on embedded coordinates and reads the visual environment.
Geo-estimation
Inferring the likely capture location of a photo from its visual content alone. No GPS. No metadata. Just the image.
You upload a photo of a street corner. It reads the road markings, the building style, the signage language, and surfaces a likely match for review: Barcelona, Catalonia, Spain.
Geolocation
The broader category. Determining where something is (a device, a person, a photo). GPS is geolocation. Cell tower triangulation is geolocation. Image-based location verification is a type of geolocation.
Your phone uses GPS satellites to geolocate you. Oceanir uses visual analysis to surface a likely capture location. Different inputs, same output type: a coordinate on a map.
Visual location verification
Review that works by analyzing what is visible in an image (architecture, terrain, vegetation, road infrastructure, signage) rather than relying on embedded data.
A photo from WhatsApp has no metadata. Visual location verification reads the stucco walls, the palm species, the dashed yellow center line. These are geographic signals.
Ground
The next step after geo-estimation, and another of Manta's arms. It takes the candidate region down to the street or the building, and only when the evidence supports it.
Geo-estimation says central Lisbon. Ground says this block of Rua Augusta, or says it cannot tell.
Reverse search
Finding visually similar images on the web. Not the same as geo-estimation. Reverse search matches images. Geo-estimation reads geography.
Google Lens finds the same building on a tourism website. That is a match. Oceanir reads a generic suburban road and surfaces a likely city for review. That is reasoning.
Landscape.
Most tools look for the same picture somewhere else. Geo-estimation reads the place in front of the camera.
Google Lens
MatchingReverse image search. Finds visually similar images on the web.
Landmarks, indexed products.
WhatsApp forwards, generic streets, screenshots.
Oceanir
ReasoningGeo-estimation. Reads the geography in the photograph itself.
Photos that were never online, evidence chains, 45.77% within 25 km on OSV-5M.
Interiors and extreme close-ups with no outdoor context.
Find where a photo was taken.
Join review teams using visual evidence verification to surface likely capture locations for human follow-up.
Try it on a photoQuestions.
Location verification is the process of reviewing where a photograph may have been captured using only the visual content of the image. It does not rely on GPS coordinates or EXIF metadata. Instead, models analyze architectural styles, vegetation, signage, terrain, and other scene-level cues to surface likely capture locations for human review.
They are closely related. Geolocation is the broader term for determining the geographic location of something (a device, a person, a photo). Location verification is a specific type of review that works purely from visual analysis of an image, without GPS, metadata, or any other data source. All location verification is geolocation in the broad sense, but not all geolocation is image-based verification.
Reverse image search finds visually similar images already indexed on the web. Location verification reviews likely capture locations from an image that may never have appeared online before. It works on novel, unseen photos by analyzing visual geography rather than finding a matching image.
Geo-estimation is inferring where a photograph was taken from what is visible in it: architecture, vegetation, road markings, signage, light. Oceanir's Geo-estimation is one of Manta's three arms. It returns ranked candidate locations, the evidence behind each, and a precision it is only allowed to claim when that evidence supports it.
On OSV-5M, the standard open-world benchmark, it places 45.77% of photographs within 25 km and 81.83% within 200 km, with a 30.5 km median error, ahead of Chipoint-2's self-reported 44.83% and 81.10%. Accuracy varies with the scene: distinctive streets resolve tighter than open country, and a photograph that only supports a region is reported as a region. A human reviewer makes the final call.
Yes. Location verification is specifically designed for images that lack metadata. Social media platforms, messaging apps, and screenshots all strip EXIF data. Location verification analyzes the visual content itself (buildings, street layouts, foliage, signage) to support review.
Journalists review the geographic origin of submitted images. Insurance teams validate claim photos. Legal teams document visual evidence. Property and corporate security teams review authorized site media. Oceanir supports these workflows without identifying people or tracking private individuals.