Facial Geometry
Facial geometry refers to the measurable shape of a face: distances, angles, and curves between points such as the eyes, nose, and chin. Computers convert these measurements into a series of numbers, which can then be used to recognize or distinguish individuals.
Every face has its own shape. The eyes are set apart at different distances, the nose varies in length, and the chin may be pointier or broader. Facial geometry is exactly that: the measurable shape of a face, expressed in distances, angles, and curves. To achieve this, a computer places many small marker points on a photo, for example at the corners of the eyes, the tip of the nose, and the edges of the mouth. From the relationships between these points, a long series of numbers emerges that is fairly typical for a given person. Software can later compare this series of numbers with others.
Why a face is treated like a password
Legally, facial geometry counts as biometric data. Biometric means: data that originates from the body itself and uniquely identifies a person. This also includes fingerprints, the iris, or the voice. Such data is sensitive because it cannot be changed. A stolen password can be replaced in two minutes. A face, however, stays with you for a lifetime.
That is why facial geometry is subject to especially strict protection in Europe. The General Data Protection Regulation, or GDPR for short, fundamentally prohibits the processing of such data and only permits it under clearly regulated exceptions. In most cases, a company needs explicit consent for this. The EU’s AI Act goes even further: it prohibits, for example, the indiscriminate scraping of facial images from the internet to build recognition databases.
The topic is also economically relevant. In the United States, the state of Illinois has its own biometrics law that requires companies to pay per affected person in case of violations. Large technology companies have therefore already settled for sums in the hundreds of millions of dollars. Anyone processing facial data therefore takes on a real financial risk.
From photo to number sequence
The first step is facial detection in the narrower sense: the software first searches the image for a face at all. It then places so-called landmarks, i.e. fixed reference points. Typically there are 68 such points, though more elaborate systems use several hundred or even thousand. From these, the program calculates ratios, such as the distance between the eyes divided by the length of the nose. Ratios are practical because they remain constant regardless of whether one stands close to or far from the camera.
Modern systems, however, hardly work with individual manual measurements anymore. Instead, a trained neural network compresses the face into a vector, i.e. a list of usually 128 to 512 numbers. This list is called a template or embedding. Two images of the same person produce vectors that lie close together. Images of different people lie far apart. The comparison then becomes pure computation involving distances.
A common misconception: the original photo cannot be recovered from the template. This is only partially true, since researchers have already reconstructed recognizable faces from such vectors. A second misconception concerns accuracy. The hit rate depends heavily on lighting, viewing angle, and image quality, and older systems were noticeably more often wrong for some population groups than for others.
From unlocking a phone to train station cameras
The most common encounter with facial geometry is unlocking a smartphone. For this, some devices project thousands of invisible infrared points onto the face and measure their distortion. This creates a three-dimensional depth map that cannot be fooled with a printed photo. The template remains in a sealed-off chip area within the device and does not travel to the cloud.
Equally widespread are automated passport checks at airports. There, a machine compares the live image with the chip photo in the passport. Banks also use this method when verifying identity via video during account opening. On social networks, the same technology underlies filters that place sunglasses precisely on the nose.
In the news, the term usually appears in debates about surveillance. This typically concerns cameras at train stations meant to match passersby against watch lists. Such live recognition in public spaces is fundamentally prohibited in the EU and is only conceivable in cases of serious crimes and with judicial authorization. So if you read about a fine against an app or a new surveillance project, facial geometry is often the technical core of the debate.