Glossary · Technology
Similarity score
A similarity score is the number a face recognition system produces when it compares two face embeddings. Higher means the faces are closer in the model's numeric space. It is a measure of resemblance under one particular model, not a probability that two photos show the same person, and the threshold that turns it into a match is a policy choice.
Updated November 6, 2026 · 1 min read
How is the score computed?
Two face embeddings are compared with a distance function, most often cosine similarity (the angle between the vectors) or Euclidean distance. The raw value is often rescaled to 0 to 100 for display. The scale is specific to the model: a 70 from one engine and a 70 from another are not comparable.
What does a threshold do?
A threshold converts the score into a yes or no. Set it high and the system misses genuine matches; set it low and it produces false positives. Engines choose thresholds to suit their use: a phone unlock tolerates very few false accepts, a web search tolerates many because a person reviews the list.
| Band | Usual meaning | What to do |
|---|---|---|
| Very high | Same photo or near-identical face | Check the page; likely the same person |
| High | Strong resemblance | Verify with name, date and context |
| Middle | Plausible look-alike | Treat as a lead only |
| Low | Weak resemblance | Usually noise |
Why is a high score not proof?
Because the index is large. In one-to-many matching against millions of faces, some unrelated face will resemble almost any probe, the doppelganger effect. Twins, relatives and low-resolution photos push scores up further. Independent testing by NIST reports error rates per threshold precisely because no single score means "same person".
Where does it come up?
In every result list of a face search engine, in access control and border systems, and in disputes about wrongful identification, where a score presented as a certainty has led to wrongful arrests documented by civil liberties groups such as the EFF.
Questions
Is a 90 score a 90% chance it is the same person?
No. It is a resemblance measure on that engine's scale. The probability of identity depends on the index size, the photo quality and the threshold, none of which the score encodes.
Why do two engines give different scores for the same pair?
Different models, different scales. Scores are only comparable within one engine.
Sources
- 1.NIST FRTE 1:N identification track, reporting on thresholds and error rates (checked December 1, 2025)
- 2.Grother, Ngan, Hanaoka, NISTIR 8280, Face Recognition Vendor Test Part 3: Demographic Effects (checked December 19, 2019)
Entry reviewed November 6, 2026. Found an error? Tell us.
Related entries
- Term Face embedding Face embedding defined: the numeric vector a network computes from a face, why same-person vectors sit close, how it is compared, and why it is biometric data.
- Term False positive False positive in face recognition defined: a reported match between different people, its relation to false negatives and thresholds, and how rates are measured.
- Term Doppelganger effect The doppelganger effect: why a face search returns look-alikes, how index size and photo quality make it worse, and how to verify a match with non-face evidence.
- Concept Accuracy, bias and testing How face recognition accuracy is measured, what NIST FRTE reports found on error rates and demographic differences, and why marketing claims cannot be checked.
Cite this entry
FaceSearch Wiki editors. "Similarity score". FaceSearch Wiki, updated November 6, 2026. https://facesearch.wiki/glossary/similarity-score/