Face recognition is a biometric identification and verification technology that locates human faces in images or video frames, extracts a compact numerical embedding of facial geometry and texture, and compares that embedding against one or more enrolled templates to establish identity. It is a specialised subfield of Computer Vision and Biometric Authentication that combines deep convolutional feature learning with metric-learning loss functions such as ArcFace and AdaFace to achieve sub-second, high-accuracy identification at scale.
Semantic Classification
Content
Compositional Relationships (Components)
SubClassOf(ai:FaceRecognition
ObjectSomeValuesFrom(ai:hasPart ai:FaceDetection))
SubClassOf(ai:FaceRecognition
ObjectSomeValuesFrom(ai:hasPart ai:FaceAlignment))
SubClassOf(ai:FaceRecognition
ObjectSomeValuesFrom(ai:hasPart ai:FeatureExtraction))
SubClassOf(ai:FaceRecognition
ObjectSomeValuesFrom(ai:hasPart ai:EmbeddingGeneration))
SubClassOf(ai:FaceRecognition
ObjectSomeValuesFrom(ai:hasPart ai:PresentationAttackDetection))
SubClassOf(ai:FaceRecognition
ObjectSomeValuesFrom(ai:hasPart ai:TemplateMatchingEngine))
SubClassOf(ai:FaceRecognition
ObjectSomeValuesFrom(ai:hasPart ai:DecisionThresholdCalibration))
Dependency Relationships
SubClassOf(ai:FaceRecognition
ObjectSomeValuesFrom(ai:requires ai:ConvolutionalNeuralNetwork))
SubClassOf(ai:FaceRecognition
ObjectSomeValuesFrom(ai:requires ai:DeepLearning))
SubClassOf(ai:FaceRecognition
ObjectSomeValuesFrom(ai:requires ai:MetricLearning))
SubClassOf(ai:FaceRecognition
ObjectSomeValuesFrom(ai:requires ai:ImageProcessing))
SubClassOf(ai:FaceRecognition
ObjectSomeValuesFrom(ai:requires ai:AnnotatedDataset))
SubClassOf(ai:FaceRecognition
ObjectSomeValuesFrom(ai:requires ai:GpuComputing))
SubClassOf(ai:FaceRecognition
ObjectSomeValuesFrom(ai:requires ai:PatternRecognition))
Capability Relationships
SubClassOf(ai:FaceRecognition
ObjectSomeValuesFrom(ai:enables ai:IdentityVerification))
SubClassOf(ai:FaceRecognition
ObjectSomeValuesFrom(ai:enables ai:DigitalIdentityVerification))
SubClassOf(ai:FaceRecognition
ObjectSomeValuesFrom(ai:enables ai:SurveillanceSystems))
SubClassOf(ai:FaceRecognition
ObjectSomeValuesFrom(ai:enables ai:EmotionRecognition))
SubClassOf(ai:FaceRecognition
ObjectSomeValuesFrom(ai:enables ai:AccessControl))
SubClassOf(ai:FaceRecognition
ObjectSomeValuesFrom(ai:enables ai:FraudDetection))
Implementation Relationships
SubClassOf(ai:FaceRecognition
ObjectSomeValuesFrom(ai:implements ai:ArcFace))
SubClassOf(ai:FaceRecognition
ObjectSomeValuesFrom(ai:implements ai:AdaFace))
SubClassOf(ai:FaceRecognition
ObjectSomeValuesFrom(ai:implements ai:CosineDistanceSimilarity))
SubClassOf(ai:FaceRecognition
ObjectSomeValuesFrom(ai:implements ai:MetricLearningObjective))
Reduction Relationships
SubClassOf(ai:FaceRecognition
ObjectSomeValuesFrom(ai:reducesTo ai:BiometricAuthentication))
SubClassOf(ai:FaceRecognition
ObjectSomeValuesFrom(ai:reducesTo ai:PatternRecognition))
SubClassOf(ai:FaceRecognition
ObjectSomeValuesFrom(ai:reducesTo ai:ImageClassification))
Support Relationships
SubClassOf(ai:FaceRecognition
ObjectSomeValuesFrom(ai:supports ai:BiometricAuthentication))
SubClassOf(ai:FaceRecognition
ObjectSomeValuesFrom(ai:supports ai:FacialActionCodingSystem))
SubClassOf(ai:FaceRecognition
ObjectSomeValuesFrom(ai:supports ai:AffectiveComputing))
SubClassOf(ai:FaceRecognition
ObjectSomeValuesFrom(ai:supports ai:DigitalIdentityFramework))
SubClassOf(ai:FaceRecognition
ObjectSomeValuesFrom(ai:supports ai:AccessControl))
SubClassOf(ai:FaceRecognition
ObjectSomeValuesFrom(ai:contrastsWith ai:IrisRecognition))
SubClassOf(ai:FaceRecognition
ObjectSomeValuesFrom(ai:contrastsWith ai:FingerprintRecognition))
SubClassOf(ai:FaceRecognition
ObjectSomeValuesFrom(ai:dependsOn ai:AnnotatedDataset))
SubClassOf(ai:FaceRecognition
ObjectSomeValuesFrom(ai:dependsOn ai:GpuComputing))
About
Face recognition is one of the most intensively studied problems in Computer Vision and Artificial Intelligence, with a research history stretching from the early 1960s experiments of Woodrow Bledsoe’s photographic ID matching through to today’s multi-billion-parameter transformer systems that exceed human performance on standard benchmarks by significant margins. Its enduring research and commercial significance arises from the combination of universality — every person has a face, making it the only truly contactless, passive, and non-intrusive biometric at scale — and difficulty: the human face is a high-dimensional, flexible surface that varies enormously with illumination angle, camera resolution, head pose, ageing, occlusion by hair or accessories, and intentional disguise, while remaining substantially similar across the millions or billions of identity pairs a practical identification system must correctly distinguish.
Face recognition has evolved from early geometric distance-based methods (Kanade, 1973; Turk and Pentland’s Eigenfaces, 1991) through appearance-based statistical models (Fisherfaces, Active Appearance Models) to the current generation of deep metric-learning systems that surpassed human-level performance on the Labelled Faces in the Wild (LFW) benchmark circa 2014. The technology is fundamentally a two-sub-problem pipeline: first, a face detector must localise and crop all face regions in a scene — modern detectors using MTCNN, RetinaFace, or YOLO-face achieve near-perfect detection under frontal and near-frontal conditions — and second, a feature extractor must map each crop to an embedding vector that is invariant to illumination, pose, age, and facial expression while remaining discriminative across millions of distinct identities. The third element is a similarity metric and decision threshold that determines whether two embeddings belong to the same person.
The distinction between face verification and face identification is fundamental to the architecture and performance evaluation of any system. Verification is a one-to-one comparison (is this probe face the same person as this reference face?) that yields a binary accept/reject decision at a pre-selected false-acceptance-rate operating point; it is the mode used in smartphone unlock, border eGate matching, and banking KYC selfie checks. Identification is a one-to-N search (who is this person, among all enrolled identities?) that yields either a ranked list of candidates or a rejection decision when similarity to all gallery members falls below a threshold; it is the mode used in law enforcement watchlist matching, airport walk-through gates, and forensic CCTV analysis. The computational and statistical challenges differ materially between these modes: verification requires high discrimination at the per-pair level; identification at scale requires that the false positive rate across the entire N×M cross-gallery comparison matrix remains manageable, imposing much stricter per-pair false-acceptance requirements as gallery size grows.
The deep learning revolution fundamentally reframed face recognition as a supervised metric-learning problem over large-scale identity datasets. Loss functions moved from vanilla softmax classification (which does not enforce tight intra-class clusters) through centre loss, ring loss, and normalisation-based approaches, to the additive angular margin family: SphereFace (2017), CosFace (2018), ArcFace (2019), CurricularFace (2020), and AdaFace (2022). ArcFace introduces an additive angular margin penalty in the normalised embedding space, providing a clear geometric interpretation: it enforces that the angle between an identity’s embedding and its class centre is smaller than the angle to any other class centre by a fixed margin — a tighter and more stable objective than multiplicative or cosine-margin alternatives. AdaFace adapts the margin dynamically based on estimated image quality, giving lower weight to unidentifiable low-quality samples during training and higher weight to hard, high-quality samples, which improves robustness under real-world surveillance conditions where image quality is highly variable.
The deep learning revolution fundamentally reframed face recognition as a supervised metric-learning problem over large-scale identity datasets. Loss functions moved from vanilla softmax classification (which does not enforce tight intra-class clusters) through centre loss, ring loss, and normalisation-based approaches, to the additive angular margin family: SphereFace (2017), CosFace (2018), ArcFace (2019), CurricularFace (2020), and AdaFace (2022). ArcFace introduces an additive angular margin penalty in the normalised embedding space, providing a clear geometric interpretation: it enforces that the angle between an identity’s embedding and its class centre is smaller than the angle to any other class centre by a fixed margin — a tighter and more stable objective than multiplicative or cosine-margin alternatives. AdaFace adapts the margin dynamically based on estimated image quality, giving lower weight to unidentifiable low-quality samples during training and higher weight to hard, high-quality samples, which improves robustness under real-world surveillance conditions where image quality is highly variable. Training on progressively larger and more diverse datasets has been a parallel driver of accuracy: from VGGFace (2 million images, 2015) through MS-Celeb-1M (10 million images, 2016) to Glint360K (17 million images, 360,000 identities, 2021) and WebFace260M (260 million images, 4 million identities, 2021), with each step producing measurable gains on hard benchmarks even with the same architecture and loss function, reflecting that face recognition performance is as much a data problem as a model problem.
The architecture of the embedding backbone has also evolved: ResNet-50 and ResNet-100 with IResNet modifications were dominant from 2018 to 2022; Vision Transformer (ViT) backbones entered the field competitively from 2021 (Face Transformer) through 2023 (TransFace, ICCV 2023) and 2025 (LVFace, ByteDance, ICCV 2025 Highlight), with LVFace achieving 97.25% TAR@FAR=1e-5 on IJB-C — a challenging cross-pose and cross-quality benchmark. Foundation model approaches using CLIP ViT backbones for face recognition have also been explored, though achieving 83% clean accuracy on FaceScrub compared to the 97%+ of specialised fine-tuned models, demonstrating that domain-specific training remains essential. The persistent accuracy gap between CLIP-style general-purpose vision models and task-specific face recognition models trained with metric learning losses is an important design lesson: pre-training on internet-scale image-text pairs optimises for semantic visual similarity, not for fine-grained identity discrimination, and domain-specific fine-tuning with an appropriate loss function remains necessary for high-stakes biometric applications.
Demographic fairness in face recognition has emerged as a research and regulatory priority following NIST FRVT findings that documented statistically significant differences in false-non-match rates (FNMR) across age, sex, and country-of-birth cohorts for most commercial algorithms tested, with some algorithms showing 10–100× higher FNMR for certain demographic pairings relative to the overall average. The causes are partly dataset composition (training sets historically over-representing younger, male, lighter-skinned subjects from North American and European image sources), partly intrinsic task difficulty differences across demographic groups (within-pair variation is higher for older subjects and for subjects photographed in variable lighting conditions), and partly architectural or loss-function interactions that amplify these imbalances. Mitigation strategies researched as of 2026 include: demographically balanced training sets (BUPT-Balanced dataset); fairness-aware loss functions that impose a ceiling on demographic performance differentials; adversarial debiasing that removes demographic information from the embedding; and independent third-party demographic evaluation — the last being mandated as a conformity-assessment element for high-risk AI systems under the EU AI Act.
Components / Architecture
The canonical production face recognition pipeline is a sequential processing chain in which each stage builds on the output of the previous, with each stage having distinct performance-limiting failure modes that must be addressed independently. Understanding the architecture at component level is essential for system integration, performance debugging, and regulatory conformity assessment.
Face Detection: The first stage localises all human face regions in the input image or video frame, producing bounding box coordinates, a confidence score, and typically five or more facial landmark locations (eye centres, nose tip, and mouth corners). Modern production detectors use anchor-free single-shot architectures (RetinaFace — the most widely used academic and open-source standard; SCRFD — the lightweight InsightFace variant optimised for mobile deployment; YOLO-Face — high-speed detection for real-time surveillance use cases) that run at tens to hundreds of frames per second on a GPU. Detection accuracy is extremely high for frontal and near-frontal faces at moderate resolution; performance degrades for extreme poses (>45° yaw), heavy occlusion (e.g., full-face mask), very small face sizes (below approximately 30×30 pixels on-sensor), and strong out-of-focus blur. Failure at the detection stage results in missed identities rather than false alarms, making missed detection (false rejection at the gallery-search level) a significant operational concern in surveillance applications with large crowds.
Face Alignment: Using the detected landmark positions, a geometric transformation (typically a similarity or affine transform) maps the detected crop to a canonical coordinate frame where the eyes, nose, and mouth corners occupy fixed positions in a 112×112-pixel (ArcFace standard) or 224×224-pixel (ViT-backbone standard) aligned crop. Alignment reduces the within-person embedding variance due to pose variation, which would otherwise inflate false rejection rates, and is particularly important for severe yaw and pitch conditions where the spatial relationship between visible features and their ground-truth positions diverges significantly. Systems that skip alignment or use low-quality landmark detectors suffer measurably on cross-pose benchmarks like IJB-C.
Feature Extraction / Backbone: The core deep neural network processes the aligned crop to produce an embedding vector. The IResNet-50 and IResNet-100 (variants of ResNet with modifications for 112×112 input and without max-pooling) were the dominant production architectures from 2018 to 2023, with the 100-layer variant achieving approximately 300 million parameters and a 512-dimensional output embedding. From 2023 to 2025, Vision Transformer (ViT) backbones (ViT-S, ViT-B, ViT-L from the DINO, DINOv2, and EVA-ViT families) entered production-quality face recognition, benefiting from large-scale vision pretraining. LVFace (ByteDance, ICCV 2025 Highlight) uses progressive cluster optimisation with a cosine-schedule training regime to stabilise ViT training on large-scale face datasets, and achieves state-of-the-art 97.25% TAR@FAR=1e-5 on IJB-C. The backbone is trained with ArcFace or AdaFace loss on datasets of 5–260 million images across 93,000–4,000,000 identities, with larger datasets consistently producing better embeddings for hard cross-quality and cross-age scenarios.
L2 Normalisation: The raw output embedding vector is unit-normalised to the surface of a hypersphere, making cosine similarity equal to the dot product between two normalised vectors. This step is standard in all ArcFace-family systems and is conceptually linked to the angular margin formulation: the margin is defined in terms of angle on the unit hypersphere, so the metric used for normalised embeddings (cosine similarity) is directly interpretable as an angle cosine.
Gallery Indexing and Similarity Scoring: For verification, the cosine similarity between a single query embedding and a single reference embedding is computed in O(d) time (d = embedding dimension = 512). For large-scale identification against a gallery of N enrolled identities, exact cosine similarity search requires O(N×d) operations per query, which is feasible for N ≤ 10,000 but prohibitive for N ≥ 1,000,000. Approximate nearest-neighbour (ANN) libraries (FAISS with HNSW or IVF-PQ indices from Meta; ScaNN from Google; Milvus for distributed deployment) compress this to sub-linear time with small accuracy trade-offs. A calibrated decision threshold converts the similarity score to an accept/reject decision at the required FAR operating point.
Presentation Attack Detection (PAD): A parallel classification module determines whether the biometric sample originates from a live human or a spoof artefact. Attack types range from 2D (printed photograph, digital replay on screen) through 3D (silicone mask, 3D-printed face) to digital injection (AI-generated deepfake image or video injected directly into the digital signal path, bypassing the camera). 2D and 3D attacks are countered by liveness signals including texture analysis (the microstructure of printed paper or silicone differs from living skin), motion-based challenges (asking the user to blink, nod, or speak), and depth/structured-light sensors (Apple Face ID TrueDepth). Digital injection attacks require anti-spoofing at the camera pipeline or API authentication layer rather than image-content analysis. ISO/IEC 30107-3 standardises the test methodology; iBeta Level 2 (testing against physical artefacts) was the dominant certification in 2024; iBeta Level 3 (introduced mid-2025, testing against AI-generated attacks) set a new upper certification tier.
Age Compensation and Quality Assessment: As the elapsed time between enrolment and verification grows, face appearance changes due to ageing (weight change, hair style, facial hair, skin condition) and the distribution shift between enrolment and probe embeddings increases, raising FRR. Quality-aware loss functions (AdaFace) address this indirectly during training by down-weighting low-quality (blurry, occluded, aged) training samples; explicit age compensation models may apply a learned transformation to the embedding to project it toward the appearance expected at the current date given the enrolled date. For high-security applications (passports valid for 10 years, driving licences valid for 10 years), age compensation is operationally necessary to maintain acceptable FRR without loosening the threshold and raising FAR.
Use Cases / Major Families
Face recognition systems divide into four broad deployment families, each characterised by distinct technical requirements, regulatory constraints, and performance trade-offs:
Cooperative, on-device verification (smartphone unlock, banking KYC selfie, physical access control): The user presents their face frontally and willingly; conditions are moderately controlled; the system performs 1:1 verification against a pre-enrolled template stored on-device or in a secure server. Apple Face ID (TrueDepth structured-light 3D depth map plus 2D IR image) achieves reliable verification in the dark and under significant disguise (glasses, hats) but uses a depth sensor unavailable in most Android devices. Android face unlock typically uses the front-facing RGB camera, with varying levels of security: high-security implementations use software-based active liveness (challenge-response blink/smile), while lower-security implementations are vulnerable to photograph attacks. Web-based remote identity proofing services (Onfido, Jumio, iProov, Facephi, Incode) perform KYC onboarding by matching a selfie video against a presented government document photograph, with ISO/IEC 30107-3 Level 2 certified liveness detection to counter deepfake injection. These systems became a primary target of AI-generated face swap attacks in 2024–2025 as diffusion-model-based face swapping reached near-photorealistic quality at consumer compute budgets, prompting iBeta to develop Level 3 PAD testing standards in mid-2025 specifically targeting generative AI spoofs.
Border control and watchlist screening (eGate biometrics, airport walkthrough, VIP access, stadium entry): 1:N identification against a known gallery at controlled operating points. IDEMIA’s real-time face recognition system, deployed in 42% of French border control lanes by March 2025, processes 60 faces per second and targets sub-500ms end-to-end latency including detection, alignment, embedding extraction, gallery search over the Schengen Information System, and decision rendering. NEC NeoFace was ranked world’s most accurate face recognition system in the April 2025 NIST FRVT update; NEC signed a deal with Japan’s Immigration Services Agency in March 2025 to deploy facial walkthrough gates at Haneda, Kansai, and Narita airports, enabling hands-free biometric passport validation for Japanese nationals and registered foreign nationals. The UK Border Agency and e-Borders programme have deployed facial biometrics at over 270 automated border control gates at UK airports, using face-to-passport-chip matching conformant with ICAO Doc 9303 Part 9 facial image data standards. Crucially, all border biometric applications involve a cooperative subject presenting their travel document, making them verification (not identification) scenarios with a tightly controlled gallery; the FAR can be set aggressively low (1 in 10,000 or stricter) and the FRR managed by requiring a retry or officer escalation.
Surveillance and public-space identification (law enforcement CCTV matching, real-time crowd screening, watchlist alerts at events): The most legally and ethically contentious deployment. In 1:N identification mode against a watchlist of hundreds or thousands of individuals, even excellent algorithms produce many false positives at population scale due to the multiplication of false acceptance rates across large numbers of comparisons per hour. The EU AI Act (in force February 2025) prohibits real-time remote biometric identification by law enforcement in publicly accessible spaces except under narrow carve-outs (imminent terrorist threat, missing child, serious crimes) that require prior judicial or independent administrative authorisation. Post-hoc remote biometric identification (matching CCTV footage after an event) is classified as high-risk AI requiring conformity assessment, registration in the EU AI Act database, and demographic bias testing. The UK, having left the EU before AI Act adoption, has no equivalent statutory framework; the Metropolitan Police and South Wales Police have operated Live Facial Recognition deployments since 2018 and 2020 respectively under a combination of common law powers, CCTV regulations, and Home Office guidance, without statutory basis. By 2025, UK police had biometrically scanned over 7 million faces. A 2026 Home Office consultation on statutory LFR governance was expected but not yet published. Clearview AI — which built a 50-billion-image database by scraping public social media without consent — was fined over EUR 110 million by five European supervisory authorities between 2022 and 2024 for GDPR violations and was ordered to delete all EU residents’ data; it continues operating in the US and non-EU jurisdictions.
Affective and behavioural analytics: Face recognition combined with Facial Action Coding System AU detection and Emotion Recognition enables engagement measurement, pain assessment, and mental-health screening. Clinical and research applications include PTSD hyperarousal detection via microexpression analysis, depression screening via longitudinal facial dynamism tracking, automated pain scoring in non-verbal patients using PSPI (Prkachin-Solomon Pain Intensity) derived from FACS AUs, and autism spectrum disorder assessment via AU timing analysis. Commercial applications include advertising attention and emotion analytics (Realeyes, Affectiva/Smart Eye, iMotions), driver monitoring systems (Seeing Machines, Smart Eye) that track AU45 blink rate and AU43 eye closure for drowsiness detection, and customer experience analytics at retail and hospitality venues. The validity and regulatory acceptability of these applications varies widely: clinical pain assessment is moving toward validation in NHS pathways, while employment emotion analytics has attracted strong ICO and EU regulatory objections.
Digital identity and financial services: The eIDAS 2.0 regulation (EU 910/2014 amended 2024) mandates that EU member states provide Digital Identity Wallets to all citizens by 2026, with high-assurance level requiring biometric identity proofing (selfie matching against passport chip data). This has driven significant commercial activity in identity verification APIs and SDKs, with face recognition as the core matching technology. UK equivalents include the UK Digital Identity and Attributes Trust Framework (DIATF, 2024), which certifies identity verification services to confidence levels, with higher levels requiring biometric document matching. In financial services, AML/KYC regulations require remote customer onboarding to include face-to-document biometric matching; the FATF guidance on digital identity has been adopted by major financial regulators globally, creating a regulatory demand floor for face recognition deployment in the financial sector.
Formal Detail: Loss Function Mathematics
The ArcFace loss is defined as:
L = -log(e^(s·cos(θ_yi + m)) / (e^(s·cos(θ_yi + m)) + Σ_{j≠yi} e^(s·cos(θ_j))))
where θ_yi is the angle between the embedding and the target class centre weight vector, m is the additive angular margin (typically 0.5 radians), and s is the feature scale (typically 64). This enforces that an identity’s embedding must be within m radians of its class prototype even after accounting for within-class variation.
AdaFace replaces the fixed margin with m(q) = -m·q + m, where q is a normalised image quality indicator derived from the norm of the pre-normalisation feature vector. Low-norm features (blurry, occluded samples) receive a near-zero effective margin, while high-norm features receive the full margin, concentrating gradient signal on informative samples. The choice of scale parameter s is critical: too small and the loss fails to discriminate between hard negatives; too large and training becomes numerically unstable. Standard practice is s = 64 for ArcFace with m = 0.5, achieving a balance that has remained largely unchanged across the major papers in the family, demonstrating the robustness of the angular margin approach to hyperparameter variation.
Benchmarks
Face recognition benchmarks have evolved from simple frontal-face test sets to challenging in-the-wild and demographic-stratified protocols:
-
LFW (Labelled Faces in the Wild, 2007): 13,233 images, 5,749 identities, 6,000 same/not-same image pairs. Saturated at ~99.8% by deep methods circa 2015; no longer a useful discriminator for state-of-the-art systems.
-
IJB-B / IJB-C (IARPA Janus Benchmark): Template-based evaluation with mixed media (still images and video frames) from unconstrained capture conditions. IJB-C (2018) has 3,531 subjects, 31,334 still images, and 117,542 video frames. TAR@FAR=1e-4 and 1e-5 are the canonical operating points. LVFace achieves 97.25% TAR@FAR=1e-5 on IJB-C (2025).
-
RFW (Racial Faces in the Wild): Demographic-balanced benchmark with four racial groups (Caucasian, Asian, Indian, African) to expose differential performance.
-
AgeDB / CALFW / CPLFW: Ageing and cross-pose/illumination benchmarks for evaluating invariance properties.
-
NIST FRVT Ongoing: The operationally relevant continuous benchmark, testing commercial and academic algorithms on mugshot (controlled), visa (controlled, cross-time), border (operational), and wild (social media) protocols; results published quarterly and used by government agencies in procurement decisions. NEC NeoFace ranked first overall in April 2025.
-
MFR-Ongoing (Masked Face Recognition, InsightFace): Online leaderboard for masked face recognition, launched during COVID-19 and maintained continuously; LVFace ranks first on the academic track as of March 2025.
Academic Context
The field’s modern foundations trace to the 2012 DeepFace paper (Yaniv Taigman, Facebook AI Research, later acquired by Meta), which introduced a nine-layer deep CNN trained on 4.4 million faces across 4,030 identities and achieved 97.35% on LFW — a stunning 27% reduction in error versus the prior state of the art — by training end-to-end on a proprietary dataset an order of magnitude larger than anything publicly available. The 2015 FaceNet paper (Florian Schroff, Google) improved further to 99.63% LFW accuracy using triplet loss training on 100–200 million face images, establishing the triplet-sampling paradigm (anchor, positive, negative) as the conceptual template from which all subsequent margin-based loss functions derived. The practical insight that set FaceNet apart was that the embedding space must be structured globally — not just locally at the sample level — and that the quality of the hard-negative mining strategy matters as much as the architecture.
The landmark LFW benchmark (Huang et al., 2007) was effectively saturated by 2015 at ~99.8% accuracy by deep methods, prompting the development of harder cross-pose (IJB-A, IJB-B, IJB-C, TinyFace) and demographic-balanced (RFW, DiveFace, BUPT-Balanced) benchmarks. The IJB-C benchmark (Maze et al., 2018) remains the dominant academic ranking benchmark as of 2026, with the reported metric being TAR at FAR=1e-5 (one false acceptance per 100,000 comparisons). The challenge of IJB-C lies in its inclusion of video frames (which may be blurry or off-angle), template-based evaluation (where a probe may contain multiple images of unknown quality), and its construction from unconstrained internet media rather than laboratory capture.
Key research groups include: MSU Computer Vision Lab (Xiaoming Liu — AdaFace, IJB-C benchmark design, NIST demographic bias analysis); Imperial College London (FACER2VM project, collaboration with Home Office and Metropolitan Police, partially occluded face recognition under adversarial conditions); University of Manchester Face Perception Laboratory (investigating human vs. machine face matching performance, individual differences in face recognition ability); Carnegie Mellon University (CMU-MultiPIE multi-pose dataset, constrained comparison analysis); NIST (FRVT ongoing vendor testing programme, demographic differential reporting); ByteDance AI Lab (LVFace, ViT-based large vision model scaling for face recognition); and the Chinese Academy of Sciences (InsightFace open-source toolkit, ArcFace training infrastructure, MFR-Ongoing leaderboard).
The NIST FRVT ongoing programme publishes continuous ranked evaluations of commercial and academic algorithms on mugshot (controlled), visa (controlled, cross-time), border (operational, mixed quality), and wild (social media images) protocols. FRVT results have documented statistically significant demographic differentials in false-non-match rates across age, sex, and country-of-birth cohorts for most commercial algorithms tested, with some algorithms showing 10–100× higher FNMR for certain demographic pairings relative to the overall average. These findings have directly influenced the EU AI Act’s classification of face recognition as a high-risk AI application requiring mandatory conformity assessment and demographic performance auditing, and have stimulated a research subfield focused on fairness-aware face recognition. Key datasets for fairness research include RFW (Racial Faces in the Wild, four racial groups), BUPT-Balanced (balanced distribution across four races), and DiveFace (six groups of combinations of race and sex from Labelled Faces in the Wild images scraped and relabelled). The convergence of these efforts represents a recognition that technical accuracy metrics and social equity requirements must both be built into the design and evaluation of deployed face recognition systems, not treated as separate concerns.
Privacy-preserving face recognition has emerged as a distinct subfield, motivated by the irreversibility of biometric data and the impossibility of “revoking” a face template once compromised. Approaches include cancelable biometrics (applying an invertible-but-secret transform to the embedding before storage, so that different transforms yield different stored templates from the same face), biometric cryptosystems (fuzzy extractors and secure sketches that derive a cryptographic key from a face embedding within match tolerance, so that authentication produces a verifiable cryptographic proof without storing the template), and homomorphic encryption (computing cosine similarity directly on encrypted embedding vectors without decrypting them at the server). These techniques are active research areas at UCL, Edinburgh, and several European academic centres, motivated in part by GDPR requirements for data minimisation and purpose limitation when processing biometric special-category data.
Current Landscape (2026)
The face recognition landscape in June 2026 reflects a technology that is simultaneously more capable, more widely deployed, and more heavily contested than at any prior point in its history. The global market is projected at approximately USD 10 billion in 2026 (MarketsandMarkets), with major verticals including airports (walkthrough biometric gates replacing human agents), smartphones (ubiquitous face unlock on mid-range and premium handsets globally), digital identity proofing (financial services KYC, government digital identity programmes), and physical access control. The market is projected to grow to USD 24 billion by 2032 at approximately 11% CAGR, though regulatory headwinds in the EU may constrain the surveillance application segment.
By June 2026, the face recognition ecosystem is characterised by five concurrent trends:
-
ViT dominance in accuracy: LVFace (ByteDance, ICCV 2025 Highlight) and similar ViT-L models trained on WebFace260M or Glint360K achieve 97%+ TAR@FAR=1e-5 on IJB-C, surpassing IResNet-100 baselines by 1–2 percentage points in the hardest protocols.
-
Deepfake injection as primary attack vector: The 2024–2025 iBeta Level 2 and emerging Level 3 PAD certifications target AI-generated face swap and diffusion-based face synthesis attacks injected into the video stream rather than presented as physical artefacts. Over 46 PAD test reports were issued to 36 vendors in 2024, a record pace.
-
Regulatory fragmentation: The EU AI Act’s Article 5 prohibition on real-time remote biometric identification in public spaces entered effect February 2025. The UK has no equivalent statute, creating regulatory arbitrage. The US has no federal law; Illinois BIPA, Texas HB 4, and city-level bans in San Francisco, Boston, and Portland constrain private sector use but not federal law enforcement.
-
On-device and federated architectures: Privacy-preserving deployments shift matching to the device (Apple Face ID, Google Pixel Face Unlock) or use secure multi-party computation to avoid transmitting raw embeddings. Federated Learning is used to update face recognition models across device fleets without centralising face images.
-
Market growth: The global facial recognition market is projected at approximately USD 10 billion in 2026, growing toward USD 24 billion by 2032 at 11% CAGR (Grand View Research), driven by airport eGate deployments, smartphone adoption in emerging markets, and financial services KYC mandates.
UK Context
The UK has a distinctive and contested position in face recognition. On the academic side, Imperial College London’s involvement in the FACER2VM project (University of Surrey, Imperial, University of Stirling, Digital Barriers, Cognitec) represents government-funded research into identification of partially occluded or masked faces, with Home Office and Metropolitan Police collaboration. The University of Manchester’s Perception, Action and Decision Making group (under Brad Duchaine’s theoretical tradition) investigates human-AI comparison in face matching. Newcastle University and University of Kent have published on individual differences in face perception. The University of Edinburgh’s informatics department contributes to privacy-preserving biometric research.
On the deployment side, the Metropolitan Police and South Wales Police have operated Live Facial Recognition (LFR) systems on a deployments-by-deployment basis since 2020, without statutory underpinning. By 2025, UK police had scanned over 7 million faces under guidance documents and common-law interpretations. Legal challenge via the Court of Appeal (Bridges v South Wales Police, 2020) ruled specific deployments unlawful for procedural reasons but did not prohibit LFR in principle, leaving a statutory vacuum that the Home Secretary acknowledged in 2025 requires a proper governance framework. The Information Commissioner’s Office (ICO) has published guidance but lacks the explicit prohibited-practice authority that the EU AI Act confers.
Northern English industrial and public safety contexts include Greater Manchester Police, West Yorkshire Police, and South Yorkshire Police, which have explored but not systematically deployed LFR, in contrast with London. Sheffield Hallam University and Leeds Beckett have contributed applied research on CCTV analytics and object detection.
Privacy and Ethics
The ethical stakes of face recognition are exceptionally high because the biometric data involved — face images linked to personal identity — is permanent, irrevocable, and collected in many contexts without the subject’s knowledge or consent. Several intersecting concerns frame the regulatory and societal debate around the technology.
First, template irreversibility: unlike a password or a cryptographic key, a face cannot be changed if a biometric database is compromised. Once an attacker obtains a face embedding and the associated identity link, they can potentially use it to spoof identity systems across any relying party using the same or a similar face recognition algorithm, with no equivalent of a “password reset” available. This motivates strong on-device storage with hardware-backed protection, template protection cryptosystems, and the on-device biometric matching architecture implemented in FIDO2.
Second, surveillance chilling effects: the awareness of being potentially recognised and tracked in public spaces is argued by civil liberties organisations to have a chilling effect on lawful assembly, political protest, and other constitutionally protected activities. The EU AI Act’s near-prohibition on real-time remote biometric identification reflects this concern at legislative level, encoding a precautionary principle: the societal risks of enabling state-level mass facial surveillance without consent are judged to outweigh the enforcement benefits except in specifically authorised circumstances. The UK’s current legal position — permitting police LFR without statutory regulation — is argued by Privacy International, Liberty, and Big Brother Watch to violate the right to private life under Article 8 ECHR and the principle of lawfulness under UK GDPR.
Third, differential accuracy and disparate impact: NIST FRVT has shown that most commercial face recognition algorithms have higher false-non-match rates (failing to match the same person twice) for women, older people, and people with darker skin tones, and higher false-match rates (incorrectly matching two different people) for those same groups in some algorithms. These differential accuracies can translate into disparate impact in practice: a watchlist system that is less accurate for a particular demographic will either generate more false alarms (leading to wrongful stops) or require a lower threshold (reducing the ability to detect wanted individuals) for that group. This has led to erroneous matches followed by wrongful arrests in several documented US cases, catalysing legislative action in cities including San Francisco, Boston, and Portland.
Fourth, data collection without consent: mass database construction through unconsented scraping (as practised by Clearview AI and others) enables the retrospective identification of individuals in historical images and videos, effectively eliminating privacy in any photo ever taken in a public space. The widespread availability of consumer-quality face recognition APIs and open-source models means this capability is available to private individuals and small organisations, not only to state actors with large technology budgets.
Future Directions (2026-2030)
-
-
Multimodal biometric fusion: Combining face with voice, gait, and iris to achieve higher accuracy and stronger liveness assurance than any single modality alone; relevant to continuous authentication in Spatial Computing environments and to adversarial conditions where a single modality is easily spoofed. Research in heterogeneous face recognition (matching face photographs across different sensor types, including visible, near-infrared, sketch, and 3D depth) enables cross-spectral identification for law enforcement and border control.
-
Privacy-preserving face recognition: Homomorphic encryption and secure multi-party computation protocols for template matching without exposing the raw embedding; active research at UCL, Edinburgh, and TU Darmstadt; targets the irreversibility problem inherent in biometric templates. Cancelable biometrics and biometric cryptosystems offer alternative architectures that preserve re-enrolment capability after compromise.
-
Demographic bias mitigation: Fairness-aware loss functions, demographically balanced training sets (BUPT-Balanced, DiveFace), adversarial debiasing, and third-party algorithmic auditing frameworks; motivated by NIST FRVT differential performance findings and EU AI Act conformity assessment requirements for high-risk systems; expected to become a procurement requirement for government contracts in EU member states from 2026.
-
Synthetic data augmentation: Using Stable Diffusion Image Model and similar diffusion-model architectures to generate demographically balanced, pose-diverse, and age-spanning synthetic training faces, reducing dependence on real-world identity-labelled datasets and improving privacy-by-design in the training pipeline. The quality of synthetic faces from 2024–2025 diffusion models is sufficient to serve as training data, though embedding spaces of models trained exclusively on synthetic data exhibit characteristic artifacts that harm performance on real-face test sets.
-
Neuromorphic and edge inference: Ultra-low-power face recognition on neuromorphic chips (Intel Loihi 2, BrainChip Akida, SpiNNaker 2) for IoT sensors and wearable devices where battery constraints preclude GPU inference; enabling always-on biometric sensing at sub-milliwatt power budgets relevant to healthcare wearables and smart home access control.
-
Regulatory harmonisation: EU–UK alignment on biometric AI governance post-Brexit; potential UK statutory framework for Live Facial Recognition in the 2027–2028 legislative cycle following the Home Secretary’s 2025 acknowledgement of the regulatory gap; development of global interoperability standards for biometric data at ICAO and ISO level.
Research & Literature
- Taigman, Y., Yang, M., Ranzato, M., & Wolf, L. (2014). DeepFace: Closing the Gap to Human-Level Performance in Face Verification. CVPR 2014.
- Schroff, F., Kalenichenko, D., & Philbin, J. (2015). FaceNet: A Unified Embedding for Face Recognition and Clustering. CVPR 2015.
- Liu, W., Wen, Y., Yu, Z., Li, M., Raj, B., & Song, L. (2017). SphereFace: Deep Hypersphere Embedding for Face Recognition. CVPR 2017.
- Wang, H., Wang, Y., Zhou, Z., Ji, X., Gong, D., Zhou, J., Li, Z., & Liu, W. (2018). CosFace: Large Margin Cosine Loss for Deep Face Recognition. CVPR 2018.
- Deng, J., Guo, J., Xue, N., & Zafeiriou, S. (2019). ArcFace: Additive Angular Margin Loss for Deep Face Recognition. CVPR 2019.
- Huang, C., Li, Y., Change Loy, C., & Tang, X. (2020). CurricularFace: Adaptive Curriculum Learning Loss for Deep Face Recognition. CVPR 2020.
- Kim, M., Jain, A.K., & Liu, X. (2022). AdaFace: Quality Adaptive Margin for Face Recognition. CVPR 2022.
- Dan, J., Liu, H., Du, J., & Deng, J. (2023). TransFace: Calibrating Transformer Training for Face Recognition from a Data-Centric Perspective. ICCV 2023.
- Kim, M., Jain, A.K., & Liu, X. (2025). 50 Years of Automated Face Recognition. arXiv 2505.24247.
- ByteDance LVFace Team. (2025). LVFace: Progressive Cluster Optimization for Large Vision Models in Face Recognition. ICCV 2025 Highlight.
- Huang, G.B., Ramesh, M., Berg, T., & Learned-Miller, E. (2007). Labeled Faces in the Wild: A Database for Studying Face Recognition in Unconstrained Environments. UMass Amherst Technical Report.
- Whitelam, C., Taborsky, E., Blanton, A., Maze, B., Adams, J., Miller, T., … & Jain, A.K. (2017). IARPA Janus Benchmark-B Face Dataset. CVPR Workshops 2017. [IJB-B]
- Maze, B., Adams, J., Duncan, J., Kalka, N., Miller, T., Charles, C., … & Phillips, J. (2018). IARPA Janus Benchmark-C: Face Dataset and Protocol. IEEE ICB 2018. [IJB-C]
- Deng, J., Guo, J., An, X., Zhu, Z., & Zafeiriou, S. (2021). Masked Face Recognition Challenge: The InsightFace Track Report. ICCV Workshop 2021.
- Wang, M., & Deng, W. (2020). Mitigating Bias in Face Recognition Using Skewness-Aware Reinforcement Learning. CVPR 2020. [Demographic bias]
- Guo, J., Zhu, X., Yang, Y., Yang, F., Lei, Z., & Li, S.Z. (2021). Towards Fast, Accurate and Stable 3D Dense Face Alignment. ECCV 2020.
- Deng, J., Guo, J., Liu, T., Gong, M., & Zafeiriou, S. (2020). Sub-center ArcFace: Boosting Face Recognition by Large-Scale Noisy Web Faces. ECCV 2020.
- NIST FRVT Ongoing. (2024). Face Recognition Vendor Testing — Ongoing Identification, Verification and Demographic Differentials. National Institute of Standards and Technology. https://www.nist.gov/programs-projects/face-recognition-vendor-testing-frvt
- ISO/IEC 19794-5:2011. Information Technology — Biometric Data Interchange Formats — Part 5: Face Image Data. International Organization for Standardization.
- ISO/IEC 30107-3:2023. Information Technology — Biometric Presentation Attack Detection — Part 3: Testing and Reporting. International Organization for Standardization.
- EU AI Act (Regulation 2024/1689). (2024). Regulation of the European Parliament and of the Council Laying Down Harmonised Rules on Artificial Intelligence. Official Journal of the European Union, L 2024/1689.
- Information Commissioner’s Office (ICO). (2023). Facial Recognition Technology in Public Places: ICO Guidance. UK ICO.
- Bridges v South Wales Police [2020] EWCA Civ 1058. Court of Appeal (Civil Division), England and Wales.
- Privacy International. (2025). Toward Regulation: Addressing the Legal Void in Facial Recognition Technology. https://privacyinternational.org/
- University of Surrey / Imperial College London / University of Stirling. (2022–2025). FACER2VM: Facial Recognition for Covered or Variously Modified faces. EPSRC/Home Office funded project.
- Phillips, P.J., Flynn, P.J., Scruggs, T., Bowyer, K.W., Chang, J., Hoffman, K., … & Worek, W. (2005). Overview of the Face Recognition Grand Challenge. CVPR 2005.
- Kortylewski, A., Egger, B., Schneider, A., Gerig, T., Morel-Forster, A., & Vetter, T. (2019). Analyzing and Improving the Image Quality of StyleGAN. CVPR 2020. [Synthetic face generation relevance to training data]
- Grand View Research. (2024). Facial Recognition Market Size, Share & Trends Analysis Report. https://www.grandviewresearch.com/industry-analysis/facial-recognition-market
Deepfake and Adversarial Threats
The rapid improvement of generative AI from 2022 to 2026 has fundamentally changed the threat model for face recognition systems, particularly in identity verification contexts where the adversary is motivated to present a convincing fake face rather than the genuine enrolled face. Deepfake face-swap technology, based on diffusion models and latent-space manipulation, can now generate photorealistic face images or video sequences of any target person from a small set of reference photographs (available from social media profiles or ID documents) in minutes on consumer hardware. The quality of these synthetic faces, when evaluated by face recognition embedding cosine similarity, is sufficient to fool systems that rely solely on 2D image content analysis. This has driven an arms race between deepfake generation and PAD (Presentation Attack Detection) systems.
The three primary injection attack vectors in biometric identity verification are: camera injection (replacing the live camera signal at the software layer with a pre-recorded or synthetic video feed); middleware injection (intercepting the video stream between camera driver and verification SDK); and API injection (sending a synthetic image directly to a cloud verification API endpoint). Camera injection is countered by secure camera attestation (cryptographic binding of the camera feed to a trusted hardware chain using device-level attestation such as Android StrongBox or iOS Secure Enclave attestation); middleware injection by OS-level sandboxing and camera API restrictions; and API injection by server-side passive liveness analysis that looks for digital artefact signatures of generated images (compression patterns inconsistent with real camera sensor noise, GAN or diffusion model generation artifacts, facial geometry inconsistencies produced by face-swap compositing). The effectiveness of these countermeasures degrades rapidly as generation technology improves, and the 2025 iBeta Level 3 PAD certification regime was specifically designed to test against state-of-the-art diffusion model deepfakes rather than the 2022-era GAN-based faces that Level 2 primarily targeted.
Adversarial attacks on face recognition go beyond deepfakes to include physical adversarial examples: carefully designed patterns printed on glasses frames, makeup, or face patches that cause misidentification or non-detection while remaining innocuous-looking to human observers. These were demonstrated in research settings (Sharif et al., 2016, Carnegie Mellon) and have been studied in physical access control contexts, though operational deployment of physical adversarial attacks against live commercial systems with PAD has not been publicly documented at scale.
Standards and Governance
Face recognition is governed by a complex and rapidly evolving interplay of technical standards, national legislation, international frameworks, and judicial precedent. At the technical layer, ISO/IEC 19794-5 specifies the biometric face image data interchange format — essentially the standard for how face photographs captured for biometric purposes (passport photos, border enrolment images) must be encoded, metadata-tagged, and stored to enable interoperability between systems from different vendors. Compliance with ISO/IEC 19794-5 is required for ICAO-compliant Machine Readable Travel Documents (ePassports) under Doc 9303, and thus for all biometric border control systems worldwide. ISO/IEC 30107-3 provides the framework for evaluating presentation attack detection systems, specifying test methodologies, reporting formats, and the classification of spoofing attack types. iBeta Quality Assurance, based in Colorado USA, is the primary accredited testing body offering PAD certification against ISO/IEC 30107-3 under NIST NVLAP accreditation.
At the regulatory layer, the EU AI Act (Regulation 2024/1689) entered into force in August 2024 and became partially applicable (including the prohibited practices provisions) in February 2025. For face recognition, the key provisions are: Article 5(1)(d) prohibiting the use of AI for real-time remote biometric identification in publicly accessible spaces by law enforcement except under judicial authorisation for specified serious crime investigations; Article 5(1)(e) prohibiting AI for untargeted scraping of facial images from internet or CCTV to build face recognition databases; and Annex III classifying post-hoc facial recognition by law enforcement as high-risk AI. The GDPR (and UK GDPR post-Brexit) classifies biometric data processed for the purpose of uniquely identifying individuals as special category personal data under Article 9, requiring an explicit legal basis for processing. NIST Special Publication 800-76 specifies biometric specifications for US PIV credentials. NIST AI 100-1 (AI Risk Management Framework, 2023) provides governance guidance for AI including face recognition systems, particularly around bias evaluation and documentation.
Relationship to Related Technologies
Face recognition occupies a specific position within a broader ecosystem of facial analysis technologies that are closely related but conceptually distinct. Facial Action Coding System (FACS) analysis shares the face detection and alignment preprocessing steps with face recognition but then diverges: face recognition extracts a global identity embedding, while FACS AU detection extracts local region-specific activation signals corresponding to muscle contractions. The two are complementary in Affective Computing pipelines: face recognition identifies who is present, while FACS-based Emotion Recognition characterises their affective state. In practice, the face recognition backbone (IResNet, ViT) is often used as a shared feature extractor whose representations are then fine-tuned separately for identity embedding and for AU occurrence prediction, because the facial features relevant to both tasks are extracted from the same low-to-mid-level convolutional feature maps.
Object Detection and Semantic Segmentation are the parent-level Computer Vision tasks from which face detection is derived — face detection is specialised object detection with a human face as the object class, and face parsing (segmenting the image into facial region components: skin, hair, eyes, lips, eyebrows) is specialised semantic segmentation. Motion Tracking is engaged when face recognition must be applied to video with multiple subjects, requiring multi-object tracking to maintain identity continuity across frames and handle entries, exits, and occlusion events. Pattern Recognition is the foundational mathematical framework from which all face recognition algorithms derive their theoretical underpinning: eigenface methods are principal components analysis applied to face image vectors, metric learning is supervised contrastive learning, and probabilistic LDA is a generative pattern recognition model for the face recognition scoring problem. Transfer Learning is routinely applied in face recognition: backbones pre-trained on large face datasets (Glint360K, WebFace260M) are fine-tuned with relatively small numbers of target-domain images when deploying to specialised populations (geriatric clinical settings with ageing subjects, infant or child biometrics, masked face recognition), where domain-specific fine-tuning closes the distribution gap between the training data and the operational deployment context.
Key Terminology
-
Face Embedding: A high-dimensional numerical vector (typically 512-dimensional, unit-normalised) that encodes the identity-discriminative features of a face crop, produced by the backbone network after the final pooling layer and before any classification head. Two embeddings from the same person should be close in cosine distance; two from different people should be far apart.
-
ArcFace / AdaFace / CosFace / SphereFace: A family of additive angular or cosine margin loss functions for training face recognition embeddings. The margin enforces that intra-class angle is smaller than inter-class angle by a fixed or adaptive amount, structuring the embedding space for discrimination.
-
FAR / FRR / EER: False Acceptance Rate (impostor accepted), False Rejection Rate (genuine rejected), Equal Error Rate (FAR = FRR). Standard operating metrics for biometric system evaluation; system designers select a FAR operating point appropriate to application risk tolerance.
-
TAR@FAR=1e-5: True Acceptance Rate at a False Acceptance Rate of one in one hundred thousand — the strictest standard IJB-C operating point, requiring that only 1 in 100,000 non-matching pairs is incorrectly accepted.
-
Verification vs. Identification: Verification (1:1) determines if a probe matches a claimed identity. Identification (1:N) searches a gallery for the identity of an unknown probe. Identification has stricter FAR requirements because the comparison set is larger.
-
Presentation Attack Detection (PAD) / Liveness Detection: The component that determines whether a biometric sample comes from a live human or a spoof artefact. Standardised by ISO/IEC 30107-3; iBeta Level 2 (2024) and Level 3 (2025) are current certification tiers, with Level 3 covering AI-generated deepfake injection attacks.
-
Template: The enrolled representation of a biometric identity — either a single face embedding or an aggregated set of embeddings from multiple enrolment images. For template-based benchmarks like IJB-B/C, a template is a set of images/frames from one identity, and the similarity between two templates is computed as the cosine similarity of their mean embeddings.
-
NIST FRVT: Face Recognition Vendor Testing — the authoritative US government continuous benchmark for face recognition algorithms, testing commercial and academic systems on controlled and operational datasets with demographic stratification.
-
IResNet: A modified ResNet variant widely used as the face recognition backbone; key modifications include removing the max-pooling layer and the stride in the first convolutional layer, retaining more spatial resolution for small 112×112 face crops.
-
Glint360K / WebFace260M / MS1MV3: Three large public face recognition training datasets. Glint360K: 17M images, 360K identities, high quality. WebFace260M: 260M images, 4M identities. MS1MV3 (cleaned MS-Celeb-1M): ~5.1M images, 93K identities.