How Accurate Is Reverse Face Search?
Reverse face search matches at approximately 95% accuracy under optimal conditions, meaning a sharp, front-facing, well-lit photo of one unobscured face. Accuracy is not a single number that holds everywhere. It falls with angle, blur, filters, and occlusion, and a confidence score measures similarity between two faces rather than the probability that a match is the right person.
Last reviewed
In short
- Approximately 95% is the figure under optimal conditions. It is not a guarantee that holds for every photo.
- A confidence score measures how similar two faces are. It is not the probability that the match is correct.
- False positives matter more than misses. Two different people scoring as a match is the error that causes harm.
- Even an accurate matcher returns wrong people at web scale, because any index holds far more strangers than it holds the person being searched for.
- Every match is a lead to verify against something else, never proof of identity.
How accurate is reverse face search?
Reverse face search reaches approximately 95% match accuracy under optimal conditions. Optimal conditions is a real constraint rather than a disclaimer: it describes one unobscured face, facing the camera, lit evenly, at enough resolution that the features are distinct. Move away from any of those and the accuracy of that individual search drops, even though the underlying model has not changed. The figure describes the system working on a good photo, not the average outcome of every photo a person happens to have.
- The number describes the matcher, not your search. Your photo decides which end of the range you land on.
- Accuracy and usefulness are different. A correct match on a page nobody can reach still tells you nothing.
- There is no accuracy figure for a face that is not indexed anywhere. That is a coverage question, not an accuracy one.
What does a confidence score actually measure?
A confidence score measures how close the uploaded face and the matched face sit in the model's numeric space. It is a similarity measurement on the engine's own scale, not a probability that the two photos show the same person. A score of 90 does not mean there is a 90% chance the match is correct. It means the two faces are highly similar by the model's measure, which is strong evidence when the photo is good and much weaker when the photo is poor.
Why does an accurate system still return the wrong person?
Because any searchable index contains vastly more strangers than it contains the specific person being looked for. A matcher that wrongly pairs one face in ten thousand sounds precise, but a search comparing against millions of indexed faces will still surface wrong people alongside any correct one. Per-comparison accuracy does not survive being multiplied by the number of comparisons a web-scale search performs. This is the reason a confidence threshold exists, and the reason the top result is not automatically the right person.
Which photo qualities change the result?
Six properties of the uploaded image account for most of the variation between a clean match and an unusable result. They compound: a small, side-on, filtered face is far worse than any one of those problems alone.
| Property | Helps accuracy | Hurts accuracy |
|---|---|---|
| Face angle | Front-facing, both eyes visible | Steep profile or head turned away |
| Resolution | The face fills much of the frame | The face is a small crop from a wide shot |
| Lighting | Even light across the whole face | Hard shadow, backlight, strong colour cast |
| Occlusion | Nothing covering the face | Sunglasses, masks, hands, heavy hair |
| Editing | The unedited original file | Beauty filters, heavy retouching, AR lenses |
| Subjects | One face in the photo | Group shots, where the wrong face may be read |
How is face recognition accuracy measured independently?
The National Institute of Standards and Technology runs the Face Recognition Vendor Test, which evaluates submitted algorithms against controlled datasets and publishes the error rates. FRVT is the reference point for how good face recognition is in general, and it separates one-to-one verification, which confirms two photos show the same person, from one-to-many identification, which is what a face search performs. FaceSearch has not published its own benchmark against those datasets, and published FRVT rates come from curated images, so they describe an upper bound rather than what any web-scale search achieves in practice.
When should a match not be relied on?
A face match is evidence to check, and it is not sufficient on its own for any decision that affects a person. The error that causes harm is the false positive: a stranger who resembles the person being searched for, surfaced with a high score and acted on as though the identification were settled.
- Do not act on a single match. Confirm it against something independent, such as a username, a mutual connection, or a second photo.
- Do not use a match to decide someone's employment, housing, or credit. Those uses are regulated in several jurisdictions.
- Do not read an empty result as proof that a person has no online presence.
- Do not assume a low score means a different person. A low score often just means a poor photo.
Sources
- Face Recognition Vendor Test (FRVT) 1:1 Verification, National Institute of Standards and Technology
- Face Recognition Vendor Test programme, National Institute of Standards and Technology