The words behind the test.
Short definitions of the terms used across the site, with a link to the page that goes further.
Updated 2026-10-18
Ages
- Chronological age
- Time since birth, the age on an ID card. Three ages.
- Apparent or perceived age
- The age people would guess from a face, or from a photo of it. What Face-Age estimates.
- Biological age
- A research estimate of how far the body’s ageing processes have gone, built from biological markers such as DNA methylation, not from a photo.
- Age group
- A range such as 30-39. Face-Age answers with one of 9 groups, because its model was trained on groups.
How a model reads a face
- Face detection
- Finding where faces are in an image. It does not say who the person is. How estimation works.
- Face recognition
- Identifying who a person is by comparing their face with others. Face-Age does not do it.
- Landmarks, face mesh
- Points placed on the face (eye corners, nose, mouth, chin). Used to align the face and to estimate head angle.
- Alignment
- Rotating and scaling the face so its landmarks land where they were in the training photos.
- Training data
- The labelled photos a model learned from. See age estimation datasets.
- ONNX
- An open file format for neural networks. Face-Age runs its age model with ONNX Runtime Web, in the browser.
- 8-bit quantisation
- Storing a model’s weights with fewer bits to make the file smaller; we measured a 0.1-point accuracy difference.
Measuring accuracy
- Exact group
- Share of faces placed in their right age group.
- Within one group
- Share of faces placed in the right group or the next one.
- Mean absolute error
- Average distance between estimate and true value, in years or, for groups, in number of groups.
- Confusion matrix
- A table crossing true groups and predicted groups. It shows which groups the model mixes up. Ours is published on the method page.
- Validation set
- Photos kept aside, never used for training, to measure accuracy fairly.
- Confidence interval
- The range in which the true value probably lies given the sample size. Small samples give wide intervals.
Age checks
- Age estimation
- Producing an estimated age or age range. Estimation vs verification.
- Age verification
- Depending on the text: a yes or no check against a threshold (NIST), or verifying the exact age (UK Online Safety Act).
- Challenge age
- A safety margin above the legal age: for 18, a threshold of 25 is common. Below it, another proof is requested.
- Liveness detection
- Checking that a real person is in front of the camera, not a photo, a video or a mask.
Photos and files
- EXIF
- Data written into a photo file by the camera: device, settings, date, sometimes GPS position. Metadata cleaner.
- GPS metadata
- Latitude and longitude stored in the file. Remove location.
- XMP, IPTC
- Other metadata blocks, used by editing software and for captions and credits.
- HEIC, HEIF
- An image format standardised as ISO/IEC 23008-12, used by iPhones; HEIC stores HEVC-compressed images. HEIC or JPEG.
- JPEG compression
- Lossy compression that removes fine detail; repeated saves and messaging apps add more.
- Pixelation
- Replacing an area with large blocks of colour. Coarse pixelation hides more than light blur, but neither is a guarantee. Blurring faces.
- INCI
- International Nomenclature of Cosmetic Ingredients: the names used in cosmetic ingredient lists. Reading an INCI list.
- Soft and hard light
- The larger a light source looks from the subject, the softer the shadows. Lighting compared.