When you use a reverse face search engine to find someone by photo, the results you see are ranked by a similarity score — a numerical value that indicates how closely each result matches your query face. This score is the output of a complex mathematical comparison between your query face print and each face print in the search database. Understanding what a similarity score means, how it is calculated, and how to interpret it is essential for making effective use of face search technology. A high similarity score does not always mean a perfect match, and a low score does not always mean no match. This guide provides a comprehensive explanation of face search similarity scores, how they are used in facesearching, and how to interpret them correctly. For the underlying technology, read our complete guide to face matching.
What Is a Similarity Score
A similarity score in face search is a numerical measure of how closely two face prints resemble each other. It is typically expressed as a value between 0 and 1, where 0 indicates no similarity and 1 indicates a perfect match (the same face). In some systems, the score is expressed as a percentage between 0% and 100%. The similarity score is calculated by comparing the face print of the query image against each face print in the search database using a distance metric. The most common metric is cosine similarity, which measures the cosine of the angle between the two face print vectors. Two face prints that point in nearly the same direction have a cosine similarity close to 1, indicating a strong match. Two face prints that point in very different directions have a cosine similarity close to 0, indicating different people. Euclidean distance — the straight-line distance between the two vectors in the embedding space — is also commonly used, with smaller distances indicating stronger matches. For more on the mathematics, read our complete guide to face embeddings.
How Similarity Scores Are Calculated
The calculation of a similarity score begins with the extraction of face prints from both the query image and the database images. Each face print is a fixed-length vector of numbers — typically 128, 256, or 512 floating-point values — that represents the unique geometric characteristics of the face. Once both face prints are available, the similarity score is computed using a distance metric. For cosine similarity, the formula is the dot product of the two vectors divided by the product of their magnitudes, resulting in a value between -1 and 1 (though in practice, face prints from the same model will produce values between 0 and 1). For Euclidean distance, the formula is the square root of the sum of squared differences between corresponding elements of the two vectors. facesearching uses optimized similarity computation algorithms that can compare a query face print against billions of database face prints in milliseconds, thanks to approximate nearest neighbor search techniques. The results are then ranked by similarity score, with the highest scores presented first.
What Scores Mean: High, Medium, and Low
Interpreting similarity scores requires understanding that there is no universal threshold that definitively separates matches from non-matches. In general, a very high similarity score — above 0.90 or 90% — strongly suggests that the two faces belong to the same person. A medium-high score — between 0.75 and 0.90 — suggests a likely match but warrants closer inspection, as variations in pose, lighting, age, or image quality can reduce the score even for the same person. A medium score — between 0.60 and 0.75 — is ambiguous; the faces may be the same person under challenging conditions, or they may be different people with similar facial features. A low score — below 0.60 — typically indicates different people, though very poor image quality can sometimes produce low scores for the same person. The key is to never rely on the similarity score alone — always review the actual images in the results and use your judgment. facesearching presents results with clickable links to the original sources so you can see the full context. For more on interpreting results, read our face search accuracy and limitations FAQ.
Confidence Thresholds and False Positives
Every face search engine uses a confidence threshold — a minimum similarity score below which results are not returned. Setting this threshold involves a trade-off. A high threshold reduces false positives (incorrectly matching different people) but may miss true matches (false negatives). A low threshold captures more true matches but increases the number of false positives. facesearching uses a carefully calibrated threshold that balances precision and recall, optimized through extensive testing on diverse datasets. The threshold is designed to return high-confidence matches while minimizing false positives. However, no threshold is perfect, and users should always verify results by examining the source images and context. The threshold may also vary depending on the search context — for example, a higher threshold might be used for identity verification, while a lower threshold might be acceptable for exploratory searches where missing a potential match is more costly than reviewing a few false positives.
Factors That Affect Similarity Scores
Several factors can influence the similarity score between two face images. Image quality is the most significant factor — blurry, low-resolution, or poorly lit images produce less reliable face prints and lower similarity scores. Pose variation — the angle at which the face is captured — can significantly reduce scores, as face prints are most accurate for front-facing images. Age differences between the query photo and the database photo can reduce scores, as facial features change over time. Facial expressions, accessories like glasses or hats, and occlusions can also affect scores. The quality of the face detection and alignment algorithms matters as well. facesearching's AI models are trained to be robust to these variations, but no system is perfect. For the best results, use a clear, front-facing photo with even lighting as your query image. For more tips on getting good results, read our step-by-step guide to reverse face search.
A similarity score is a probability, not a certainty. It tells you how likely two faces are the same person, but it is not a guarantee. Always verify results by looking at the actual images, not just the numbers.
Using Similarity Scores Effectively
To use similarity scores effectively in your face search workflow, follow these best practices. First, always start with the best quality query photo you can obtain — clear, front-facing, well-lit. Second, review the top results first, as they have the highest similarity scores and are most likely to be correct matches. Third, examine the actual images in the results, not just the scores — two faces with a moderate similarity score may still be the same person under challenging conditions. Fourth, consider the context: does the person in the result appear in a location, profession, or social circle that is consistent with what you know about the person you are searching for? Fifth, if you get no high-confidence matches, try a different photo of the same person. facesearching's results include clickable links to source profiles and articles, so you can explore the full context of each match. Ready to try a face search and see similarity scores in action? Visit facesearching today.