Spetslab: the accuracy of face identification systems cannot be measured by a single number

Specialab published an analysis explaining why the question of "accuracy" in facial identification systems is fundamentally misframed. The performance of such systems depends not on one magic number, but on face angle, use case scenario, threshold settings, and how the image database is structured.

Why There Is No Single Number

The main point of the article is simple: a facial identification system does not answer the question "did it recognize the person with absolute accuracy." It always searches for the most similar face in the database and compares the current frame with the samples it already has. While a person looks directly at the camera, the task is relatively straightforward. But the more the face is turned, the worse the lighting, sharpness, and shooting angle, the more the number of similar candidates grows. This is why the same algorithm can show almost error-free results at an access point and noticeably more often make mistakes in a crowd of people.

"In one configuration errors are eliminated, but the face looks

strictly frontal."

The author emphasizes that users often expect from biometrics a universal rating, similar to household appliances or speed tests. In reality, no such rating exists because the system always balances between strictness and flexibility. If you set a maximum strict threshold, only good frontal frames will reliably pass. If you loosen the settings, you can find a person in more difficult poses, but at the same time the risk of false matches grows. It is this compromise that determines the actual effectiveness of the solution.

Access Control Systems and Search

The article discusses two almost opposite scenarios. The first is an access control system (ACCS), when a camera is installed at a pre-prepared checkpoint and the person understands they need to look at the lens. Here you can demand almost perfect frontal view, set a high matching threshold, and get very stable operation. The second scenario is searching for a specific person through a video stream, for example in an archive or in a crowd. There people do not look at the camera on command, hold a phone to their face, turn away, get into the frame from above or from the side, and therefore the system is forced to work more flexibly.

• In access control systems, the camera is positioned to capture a frontal shot.
• In crowd search, the system works with poor angles and random poses.
• Strict settings reduce errors but miss difficult frames.
• Flexible settings increase the chance of finding the right person but produce more false alarms.

Hence the conclusion: demanding a single "accuracy" number for both modes is meaningless. For an access point, it is important to almost never let strangers in and consistently recognize known people in controlled conditions. For search, it is more important to reduce the operator's manual review workload and quickly show a limited set of the most similar faces. This is no longer one and the same task, but two different usage models with different error tolerances.

How Errors Are Reduced

Specialab proposes to overcome the basic limitation not only through settings but also through the structure of the database itself. If you need to track your own employees or regular visitors in different conditions, you should add to the database not one frontal photo but several images with different head rotations. Then even a severely distorted frame will be compared not with an "ideal passport face" but with a pose-similar sample of the same person. This approach is especially useful for tracking movements throughout a facility, where it is important not to confuse your people between cameras.

The article also draws an important distinction between simple face recognition and personal identification. A face detector is needed to find a face in a frame at all, extract it from a multi-hour archive, and remove duplicates. The author gives a telling example: reviewing a week of recordings from 16 cameras could take 2,688 hours, and a face detector reduces the search to individual frames with people. However, in complex scenarios — at night, out of focus, in black and white video — a human still sometimes recognizes a familiar person better than an algorithm because they rely not only on facial metrics but also on the overall visual context.

What This Means

For business and security services, this is a good guideline: facial identification systems should be evaluated not by abstract "accuracy" but by specific scenario, shooting conditions, and the cost of error. If the task is formulated correctly, biometrics can significantly reduce manual labor and speed up search. If incorrectly formulated — even a strong algorithm will seem weak simply because it is expected to do the impossible.