A facial recognition camera matches a live face in two separate stages. First a face detection system finds each face in the frame and returns a usable crop. Then a recognition model turns that crop into a numerical template and compares it with templates stored when people were enrolled, producing a similarity score that a threshold turns into accept, reject or a ranked list of candidates. Each stage has its own technical demands and its own points of breakdown. Whether the deployment checks one claimed identity or searches everyone it has enrolled shapes its accuracy, where the matching should run and what the law allows, more than any other choice.
How a face detection system feeds the match
Detection answers “is there a face in this image, and where?”, while recognition answers “whose face is this?”. Treating them as one capability leads to unrealistic accuracy expectations and to cameras specified for the wrong job. Everything downstream (alignment, embedding, the gallery match) depends on whether the detector returned a usable crop in time, which is why real-time face detection under live-video constraints is the stage most teams underestimate. Detection is also the prerequisite for landmark localisation and alignment, and so for assessing most quality criteria.
Once a face is cropped and aligned, a convolutional network converts it into a face embedding, a single vector whose most common sizes are 128 and 512 values. That vector is compared with stored ones by Euclidean distance, by a trained classifier, or by voting across several stored images of the same person. A classifier has to be retrained whenever a person is added or removed. That makes distance matching against stored templates the practical choice for a gallery whose membership keeps changing.
Amazon Rekognition’s documentation shows the same shape from the service side: faces are stored in server-side “collections” as face vectors, a mathematical representation rather than an image, and the IndexFaces operation detects faces, extracts the vectors and stores them. The documented access-control example authenticates an employee when the similarity score is sufficiently high, for instance 99%. For the stage-by-stage mechanics (detector choice, alignment, embedding models), see how face recognition works in production; here we stay with what changes once a camera goes into a real building.
Does your camera check one claimed identity or search everyone enrolled?
The EU AI Act defines biometric verification as the automated, one-to-one verification of a person’s identity by comparing their biometric data with data they previously provided. Biometric identification, by contrast, establishes who someone is by comparing their data with that of individuals stored in a database. In practice a 1:1 system returns accept or reject against the identity someone claims, and a 1:N system returns a ranking across everyone enrolled.
The UK Information Commissioner’s Office draws the line in operational terms. In one-to-one uses the individual participates directly and knows why their data is used, while live facial recognition is directed at everyone in an area and can capture the biometric data of all passers-by automatically. A door terminal that a staff member walks up to and a ceiling camera scanning a shop floor against a watchlist are different products with different risk profiles, even if they run the same model.
The survey 50 Years of Automated Face Recognition (arXiv, 2025) reports that the leading algorithm in the latest NIST FRTE 1:N benchmark reached a false negative identification rate of 0.15% at a false positive identification rate of 0.001, on a gallery of over 10 million identities. NIST’s 1:N evaluation primarily uses standardised portrait images. That figure therefore describes standardised portrait images, not a person walking past a camera at an angle. The same survey shows rank-1 accuracy on the IJB-S surveillance protocol falling from 62.0% with 202 enrolled identities to 48.4% with 10,202, and the reasons that gap opens in the field are covered in our piece on CCTV face recognition failure in production.
What enrolment involves, and how photo quality sets the ceiling on accuracy
A modern face recognition system runs in two phases: enrolment, which captures and stores facial data as a baseline, and recognition, which uses that baseline to confirm or determine identity. Enrolment is the first of those two phases. In practice it decides most of what the matcher can later achieve.
NIST’s quality assessment concept describes the core mechanism: a scalar quality value lets a system reject an image that is too poor and trigger a new capture, and the check can run in the camera, a client computer or a remote server. NIST notes that this is most useful at initial enrolment, when no earlier reference image of the person exists. The same document advises that a recognition application should generally enrol all of the images it collected of a person rather than keep only the best one.
Vendor guidance turns that into concrete parameters. Amazon recommends indexing at least five images per person: straight on, turned left and right at a yaw of 45 degrees or less, tilted down at a pitch of 30 degrees or less, and tilted up at a pitch of 45 degrees or less. Rekognition lets up to 100 face vectors be aggregated into one user vector, and searching against user vectors can significantly improve accuracy compared with searching individual face vectors. Amazon also advises keeping recent images indexed, which matches the Council of Europe’s point that reliability should be recorded over time and photos renewed when it deteriorates, for example as faces age.
The State of the Art of Quality Assessment of Facial Images survey (arXiv, 2022) gives an example where a recognition algorithm starting at an error rate of around 4% had that error nearly halved once roughly 5% of the worst-quality images were rejected using MagFace quality scores. That result holds for one algorithm and test set; the error-versus-reject curve behind it plots error against the fraction of images rejected for poor quality.
A Study on the Impact of Face Image Quality on Face Recognition in the Wild (arXiv, 2023) found that matching middle-quality probes against a high-quality gallery performed much better than matching low-quality probes, and attributed it to the larger quality difference. Our reading: a passport-grade enrolment photo does not rescue a dim, angled capture at the door, so enrol at least some images under the conditions of the camera that will do the matching.
ISO/IEC 19794-5:2011 sets scene, photographic and digital-image criteria for face images, and the ISO/IEC 29794-5 revision defines how to measure deviation from the ISO/IEC 39794-5 requirements. Germany’s BSI publishes OFIQ, an open-source reference implementation of ISO/IEC 29794-5 that runs detection, 98-point landmark estimation, alignment and occlusion segmentation before scoring quality.
Enrolment checklist before the first match runs
- A quality score with a rejection threshold is applied at capture, and rejected images trigger a recapture.
- Several images per person are enrolled, covering the recommended yaw and pitch range, not a single “best” shot.
- Watchlist or gallery images are quality-checked to prevent false matches, since low-quality images increase errors.
- At least part of each person’s enrolment is captured on the deployment camera, to narrow the quality gap with live probes.
- Recognition reliability is dated and recorded, with a schedule for renewing enrolment photos.
- Enrolment images come from sources with a lawful basis for biometric use, not from scraped or repurposed footage.
Rite Aid: low-quality enrolment images and a five-year ban
Example: the US Federal Trade Commission’s December 2023 action against Rite Aid. According to the FTC, the retailer used facial recognition from 2012 to 2020 to identify customers who may have been engaged in shoplifting, collected tens of thousands of images, many of them low-quality and taken from security cameras, employee phone cameras and even news stories, and the system generated thousands of false-positive matches. The FTC says consumers, particularly women and people of colour, were erroneously accused by employees because the system flagged them as matching someone previously identified as a shoplifter. Rite Aid was banned from using facial recognition for surveillance for five years.
A 1:N watchlist built from uncontrolled images, matched against uncontrolled CCTV captures and acted on by staff, combines every weak point described above.
Recognition camera or CCTV with face detection: how to tell them apart
Recognition goes a step beyond detection: it answers whose face it is, while detection only finds faces. Axis states of its AXIS Face Detector that only the face is detected and no further analysis is performed. The application runs as edge analytics on the uncompressed video feed without a dedicated server, works best in well-lit areas with good contrast, requires 0 or 180 degree camera placement, and detects only upright, frontal faces.
A recognition system adds what a detection-only camera lacks: enrolled templates, an embedding model and a matcher that compares similarity scores against a threshold. When a quotation says “face recognition camera”, ask where the gallery lives and which component computes the match. Resolution, lighting and mounting choices for the capture side are covered in face detection camera specs.
Laboratory demonstrations assume frontal pose, controlled lighting, high resolution and cooperative subjects, while CCTV delivers faces at distance, at angle, under poor lighting, partly occluded and in motion. Our position is that detection and recognition only look like solved problems in the demo: an off-the-shelf model that works in a controlled environment fails at production edge cases, lighting and throughput, although for a proof of concept it may be entirely appropriate. Whether to start from open-source frameworks, cloud APIs or proprietary SDKs depends on deployment constraints, privacy, accuracy and cost at the expected throughput, a choice laid out in our guide to open-source versus commercial facial detection software.
Where the face match runs: trade-offs between camera, edge box and central server
In an edge-centric architecture, detection, liveness, template extraction and matching all run on the device, which keeps working offline and authenticates quickly at busy doors. A server-centric architecture sends images or features to a central server that acts as the single source of truth, applies revocations instantly, scales to large biometric databases across sites and can run heavier models for accuracy. That comparison comes from an access-control vendor’s analysis published by SecurityBrief Asia, which concludes there is no one-size-fits-all architecture.
Koubaa et al. (CISTER), comparing cloud, edge and hybrid deployments of an MTCNN detector with a FaceNet recogniser, summarise the trade-offs as follows:
| Characteristic | Cloud (server) | Edge (camera or edge box) | Hybrid (detect on edge, match centrally) |
|---|---|---|---|
| Communication delay | High | Low | Low |
| Processing delay | Low | High | Medium |
| Energy | High on cloud, low on edge | High on edge, low on cloud | Balanced |
| Scalability | Limited | High | High |
| Privacy | Low | High | Medium |
In the hybrid design the edge device does capture, detection and tracking, and sends only the cropped face, which is much smaller than the full frame, for recognition elsewhere. Privacy sits in between because face data still crosses the network. The same study measured cloud GPUs running the pipeline 2x to 7.7x faster on average than Jetson edge boards. TensorRT accelerated MTCNN by 6x to 11x and FaceNet by 51% to 76% over TensorFlow, at around 40% more energy, the kind of gain explained in our piece on ML compilers for face recognition inference. Resolution trades directly against coverage: moving from 480p to HD cut speed by 2.4x on average but raised the number of detected faces by 35% to 83%.
Low-cost camera silicon offers substantially constrained compute and memory bandwidth. A 2026 Scientific Reports benchmark of Rockchip SoCs found that for surveillance at 25 to 30 FPS on 720p video, only Nano and Small YOLO models on the RK3588 met real-time constraints, while the entry-level RV1106 fell short with every model tested. Latency tracked accuracy and architecture far more than FLOPs, so the authors advise benchmarking on the target SoC rather than relying on theoretical indicators or GPU measurements. Those were YOLO object detectors, not face models, but the lesson about measuring on the actual chip carries over.
The European Data Protection Board says that in a controlled environment, templates should be stored on a device the user keeps (a smartphone or ID card) or, where objectively needed, in an encrypted central database whose key is held solely by the person. We would settle the matching mode and gallery size first, then the revocation and privacy requirements, and only then pick hardware and measure it under load. The general reasoning on splitting vision workloads is in cloud vs edge workload split.
What GDPR, Illinois BIPA and the EU AI Act allow a commercial deployment to do
GDPR and UK GDPR
GDPR Article 9(1) prohibits processing biometric data for the purpose of uniquely identifying a natural person, unless an exception applies, the first being explicit consent for specified purposes. The EDPB’s 2019 video-device guidelines state that biometric recognition installed by private entities for their own purposes, whether marketing, statistics or security, will in most cases require explicit consent from all data subjects. Their examples are practical. In the EDPB’s office example, access is triggered by the person, for instance by pushing a button, and a non-biometric alternative such as a badge must be offered. A shop using facial recognition for tailored advertising is unlawful if it captures visitors who have not consented, even if their templates are deleted within the shortest possible period. Once a match or no-match result is obtained, the templates made on the fly must be deleted immediately.
In the UK, the ICO’s 2021 opinion says live facial recognition for automatic, indiscriminate collection in public places faces a high bar to be lawful, and that consent is unlikely to be an appropriate lawful basis there. The opinion explicitly excludes one-to-one uses such as access control with prior enrolment. It also found controllers presenting vendor accuracy rates without understanding their provenance or suitability to their own use case.
Council of Europe Convention 108 guidelines
The Convention 108 guidelines go further for private operators: private entities should not deploy facial recognition in uncontrolled environments such as shopping centres to identify persons of interest, for marketing or for private security. Under those guidelines, consent-based private use belongs in controlled environments for verification, authentication or categorisation, and an impact assessment is expected before processing.
Illinois BIPA
The Illinois Biometric Information Privacy Act requires private entities to obtain written informed consent before collecting face geometry scans, with written notice of the purpose and how long the data will be kept. Entities must publish a retention schedule and destroy biometric data when the purpose is satisfied or within three years of the person’s last interaction, whichever comes first, and may not sell or otherwise profit from it. Liquidated damages are $1,000 per negligent violation and $5,000 per intentional or reckless one, or actual damages if greater. A 2024 amendment treats repeated collection of the same identifier from the same person by the same method as a single violation. One engineering-relevant line: in Doe v. Apple (S.D. Ill., 30 September 2026) a federal district court held that a face crop that only locates a face, without measuring its geometry, is not a biometric identifier. Under that ruling, a detection-only crop fell outside the statute while a face-geometry scan stays covered.
EU AI Act
Article 5 of the EU AI Act prohibits AI systems that create or expand facial recognition databases through untargeted scraping of facial images from the internet or CCTV footage. Its ban on real-time remote biometric identification in publicly accessible spaces applies to law enforcement use. The narrow exceptions are targeted searches for victims and missing persons, an imminent threat to life or of terrorism, and locating suspects of serious offences. Even then, use must only confirm the identity of the specifically targeted individual, weigh the seriousness and scale of harm against the consequences for everyone’s rights, and respect temporal, geographic and personal limits set by national law. It also requires a completed fundamental rights impact assessment and registration in the EU database. No decision with an adverse legal effect may rest solely on the system’s output. Article 3 defines a remote biometric identification system as one that identifies people without their active involvement, typically at a distance, and “real-time” as capture, comparison and identification without significant delay. A commercial operator should not read the law-enforcement scope of Article 5 as permission: the Act’s other obligations, and GDPR, still apply.
Quick comparison
| Rule | What it targets | What a commercial deployment must do |
|---|---|---|
| GDPR Art. 9 | Biometric data for unique identification | Find an exception, usually explicit consent for private use |
| UK GDPR (ICO opinion) | Live facial recognition in public places | Clear a high bar for lawfulness |
| Convention 108 guidelines | Facial recognition by private entities | Keep consent-based private use to controlled environments |
| Illinois BIPA | Scans of face geometry | Written notice and consent, retention policy, no sale |
| EU AI Act Art. 5 | Scraped face databases; real-time public RBI by law enforcement | Do not create or expand face databases by untargeted scraping of internet or CCTV images |
Who can build a facial recognition system that holds up outside the demo?
Through its Computer Vision R&D practice, TechnoLynx can build and tune face detection and recognition pipelines for commercial cameras. That work can include enrolment quality gates, threshold calibration on footage from your own cameras, and porting and optimising the models for the edge device, edge box or server the deployment needs.
If the need is a simple 1:1 door terminal with consenting staff, a packaged access-control product may be enough. A custom build makes sense for 1:N galleries, poor capture conditions or hardware constraints the packaged products do not meet.
FAQ
Is there such a thing as face recognition?
Yes. The Council of Europe defines facial recognition as the automatic processing of digital images containing faces to identify or verify individuals using face templates.
What are the five facial features?
Common face detectors such as MTCNN, RetinaFace, BlazeFace and YOLO5Face output just five or six landmarks per face. These are typically points on the eyes, nose and mouth, the key features alignment uses to position faces consistently.
Sources
- 50 Years of Automated Face Recognition, arxiv.org
- State of the Art of Quality Assessment of Facial Images, arxiv.org
- Cloud versus Edge Deployment Strategies of Real-Time Face Recognition Inference (Koubaa et al.), cister.isep.ipp.pt
- Searching faces in a collection - Amazon Rekognition, docs.aws.amazon.com
- EU AI Act Article 3: Definitions, artificialintelligenceact.eu
- ICO Opinion: The use of live facial recognition technology in public places (2021), ico.org.uk
- Face Recognition Vendor Test (FRVT), nist.gov
- Face Recognition Vendor Test Ongoing Face Recognition Quality Assessment, nist.gov
- Recommendations for searching faces in a collection - Amazon Rekognition, docs.aws.amazon.com
- Guidelines on facial recognition Convention 108 Consultative Committee of, rm.coe.int
- A Study on the Impact of Face Image Quality on Face Recognition in the Wild, arxiv.org
- Open Source Face Image Quality (OFIQ) overview – BSI, bsi.bund.de
- Rite Aid Banned from Using AI Facial Recognition (FTC press release), ftc.gov
- AXIS Face Detector datasheet, axis.com
- Edge or server? Choosing the best facial recognition architecture for access control systems, securitybrief.asia
- Object detection on low-compute edge SoCs: a reproducible benchmark and deployment guidelines, nature.com
- EDPB Guidelines 3/2019 on processing of personal data through video devices, edpb.europa.eu
- Art. 9 GDPR, gdpr-info.eu
- Illinois Biometric Privacy Laws: Collection, Consent & Penalties (2026), recordinglaw.com
- AI Act Article 5: Prohibited AI practices (European Commission AI Act Service Desk), ai-act-service-desk.ec.europa.eu
