Wiki
Advanced14 min read

Face and gait recognition

Two biometrics from video: face recognition by embedding faces into a metric space, and gait recognition from the way a silhouette moves.

Biometrics asks not "what is in this video?" but "who is this?". Two modalities dominate the non-invasive end of the spectrum. Face recognition reads the geometry and texture of a face; it is accurate, needs cooperation, and demands a good view. Gait recognition reads the way a person walks; it works at a distance and in low resolution, but needs several steps and is easy to change.

They are not competitors so much as different operating points, and comparing them clarifies what a biometric actually requires.

Verification, identification, recognition

Keep three tasks apart. Verification asks "is this the claimed person?" and is a one-to-one yes/no at a threshold. Identification asks "who is this?" against a gallery and is one-to-many. Recognition is the informal umbrella. Confusing verification accuracy with closed-set identification accuracy is a recurring source of misleading claims.

Adjust the operating conditions and watch the two modalities cross the same decision threshold.

Tune distance, face resolution and available frames — see which biometric clears the same 0.5 decision threshold.

face recognition0.362 · no match
gait recognition0.601 · match

At these settings the more reliable modality is the gait. — face fails on range/resolution while gait still has enough silhouette signal.

propertyfacegait
cooperationrequired (look at camera)not required
rangemetrestens of metres
illuminationneeds visible lightworks in darkness (IR)
resolutionhigh (face region)low (whole silhouette)
temporal inputsingle frameseveral gait cycles
privacybiometric face datasilhouette, less identifying

Illustrative score model: exponential range falloff, sigmoid resolution quality and an exposure curve for the number of gait cycles. Real systems report EER/ROC curves on datasets like LFW or CASIA-B; the qualitative trade-offs in the table are the reliable part.

Face recognition as metric learning

Modern face recognition is not template matching; it is embedding. A network maps an aligned face to a vector such that distance encodes identity: photos of the same person are close, photos of different people are far. Training uses a margin-based loss so that, for an anchor aa, a positive pp (same identity) and a negative nn (different),

∥f(a)−f(p)∥2+m  <  ∥f(a)−f(n)∥2.\lVert f(a) - f(p) \rVert^2 + m \;<\; \lVert f(a) - f(n) \rVert^2 .

ArcFace sharpens this by adding an angular margin on a hypersphere and normalising both the embedding and the class weights. At test time, verification compares the cosine similarity of two embeddings to a threshold calibrated from a validation set.

The earlier eigenfaces approach — project a face onto the leading principal components of a face image set and match in that subspace — is the linear ancestor of the same idea, and remains the clearest illustration of why a low-dimensional representation of identity is even possible.

Gait recognition from silhouettes

Gait treats the body as a moving shape. A common representation is the gait energy image (GEI): segment the person over a gait cycle, extract the silhouette each frame, and average them over the cycle. The result is a single image that encodes the characteristic shape and its motion, which a CNN then embeds just as it would a face. Because it uses the whole-body silhouette, gait survives low resolution and distance; because it depends on pose and motion, it degrades with load, footwear, injury, and viewpoint.

Measuring a biometric

A biometric is characterised by the ROC and, more compactly, the equal error rate (EER): the point where false-accept and false-reject rates are equal. Accuracy is quoted only alongside the operating point and the dataset — and for faces, alongside the demographic slice, because error rates are known to vary substantially across populations.

Biometrics raise real ethical stakes

Face recognition is a surveillance technology as much as a classifier. False matches are not abstract; they have led to wrongful arrests. Deployment choices about consent, retention, purpose limitation and demographic fairness are part of the engineering, not optional add-ons, and many jurisdictions now regulate them specifically.

Illustrative vs real

The widget's score curves are illustrative formulas, not measured performance. Real numbers come from benchmarks such as LFW, IJB-C and MegaFace for faces and CASIA-B or OUMVLP for gait. What is genuine here is the decision-threshold framing and the qualitative robustness trade-off: faces need resolution and cooperation, gait needs several cycles and tolerates distance.

Check yourself

Eduspheria wiki · Applied AI, Video analytics

0 / 5 answered

  1. 1At a given threshold a biometric system has false-accept rate 0.01 and false-reject rate 0.04. What is the sum of the two error rates?
    Numeric answer
  2. 2What does an embedding-based face recogniser compare at test time?
    Multiple choice
  3. 3Gait recognition typically needs several frames spanning at least one gait cycle.
    True / false
  4. 4Which single-image gait representation averages aligned silhouettes over a gait cycle?
    Short answer
  5. 5Which task compares a query against a gallery of many identities?
    Multiple choice

From the mid-term paper

Modeled on NITJ AI-604, Mid-Term March 2025

0 / 5 answered

  1. 1How many dimensions are in the standard SIFT keypoint descriptor vector?
    Numeric answer
  2. 2In AdaBoost, if a weak classifier has weighted error 0.1, what is its weight alpha = 0.5 times ln((1 - error) / error)? Round to two decimals.
    Numeric answer
  3. 3Which corner detector scores a point by the smaller eigenvalue of the second-moment matrix?
    Multiple choice
  4. 4Which motion-detection method subtracts a learned reference frame from each new frame to expose foreground objects?
    Short answer
  5. 5Which edge model describes a gradual intensity change caused by blur or motion?
    Multiple choice

Where next: leaving pixels behind, the next chapter models people as nodes in a network and asks who is central, who influences whom, and how facts connect.