Face recognition benchmarks and datasets, and what each one measures

Face verification metrics explains true accept rate at a fixed false accept rate, the difference between 1:1 verification and 1:N identification, and how to read a face recognition results table.

NIST's evaluations explains how the NIST FRTE series tests submitted algorithms on sequestered images, how 1:1 and 1:N accuracy is reported, and how to read a results page without over-claiming.

Demographic differentials sets out what the NIST demographic effects evaluation (NISTIR 8280) measured about false match and false non-match rates across groups, and what the numbers do not say.

WebFace260M and WebFace42M explains what WebFace260M and WebFace42M are, how the cleaned set was made, what the research licence covered, and why nothing is downloadable here.

The ICCV 2021 challenge describes the ICCV 2021 masked face recognition challenge from its public report: test sets, time limit, scoring, and where the results live now.

From face recognition to deepfake detection follows one pipeline from "who is this?" to "is this real?", and shows which benchmark habits carry across and which do not.

Liveness and injection attacks covers how identity checks are fooled at the camera or behind it, how presentation attack detection is tested under ISO/IEC 30107-3, and where a deepfake detector fits.

Gender Shades sets out what Buolamwini and Gebru measured in 2018, on which faces, with which metric, and why it is a gender classification result rather than a face recognition one.

How face recognition works walks the four steps every modern system runs, detect, align, embed and compare, with the paper and the published number behind each one.

Face recognition technology explains the parts of a face recognition system, the web-scale training sets behind modern models and the evaluations that score them.

FMR, FNMR, FPIR, FNIR defines the false match rate, false non-match rate, false positive identification rate and false negative identification rate in NIST's own words, with the threshold that links them.

NISTIR 8280 explained walks through the NIST demographic effects report section by section: its datasets, its metrics, its four results chapters and the 17 annexes behind them.

AI face recognition traces how learned embeddings and web-scale training sets saturated LFW and moved evaluation to TAR at FAR on IJB-C and to NIST operational tests.

Face recognition bias separates the two studies most often cited for face recognition bias, says what each measured, what causes are supported, and what a bias claim must specify.

System components walks through the five components of a face recognition system in plain language, with the error each one can introduce and who measures it.

Face recognition accuracy explains why face recognition accuracy is quoted as an error rate at a fixed operating point on a named dataset, and what NIST found when it measured 189 algorithms that way.

LFW dataset gives the Labeled Faces in the Wild counts, the verification protocol they feed, the three published accuracies that took it to 99.83 percent, and what replaced it.

1:1 versus 1:N separates verification from identification, follows one face through each branch, and shows how NIST reports the two with different error rates and thresholds.

MegaFace records what the MegaFace benchmark measured with a million distractors, how far LFW-saturated algorithms fell, and the University of Washington notice that decommissioned it.

ArcFace explains the additive angular margin loss, what a results table means by an ArcFace backbone, and the verification figures the CVPR 2019 paper reported on each named set.

TSA face recognition describes the TSA checkpoint face comparison program as TSA and DHS documents state it: the 1:1 match, the optional photo, the deletion rules and the 1:n program for trusted travellers.

Face recognition at airports separates the two US airport face recognition programs, CBP biometric entry and exit and TSA identity verification, by what each compares a face against and which privacy assessment covers it.

InsightFace describes the open-source toolkit behind ArcFace and RetinaFace, its licence split between code and data, and the training and test sets its dataset README lists with their identity counts.

Clearview AI rulings sets out each regulator's Clearview AI decision with its date, amount and order, read from the decision pages, and follows the UK penalty through two tribunals.

Benchmarks list puts five face recognition benchmarks in one table, with each one's protocol, metric, size and distribution status, then explains what each measures.

What is face recognition defines face recognition, separates it from detection, attribute classification and deepfake detection, and names the two tasks and the error rates that score each.

Biometric face recognition explains face recognition as a biometric system, the ISO/IEC 19795-1 testing framework, the four error rates and how NIST applies them at fixed operating points.

Face recognition test sets out what a face recognition test fixes before any score is quoted, from the pair list and threshold policy to the metric and the report, with the NIST and ISO rules.

WebFace260M paper walks through the CVPR 2021 WebFace260M paper: how 260 million faces were collected, how CAST cleaned them to 42 million, what the FRUITS time budgets test and what the tables show.

Law enforcement use lists what official audits and court findings have recorded about police and federal use of face recognition, with dates, and what those documents did not measure.

Face recognition in China reads China's Personal Information Protection Law, the 2021 Supreme People's Court interpretation and the 2025 facial recognition measures as their official texts state them, with deployments only where a document records them.

TAR at FAR works through one published IJB-C verification row, 97.2 percent TAR at FAR 1e-4, in genuine and impostor pairs, and shows how the same model's TAR moves when the FAR moves.

Open-source face recognition compares the three open toolkits most papers and libraries build on, the licences that split code from models, and the published scores that change with the training set behind each model.

Facial recognition privacy puts the laws, regulator decisions and dataset withdrawals that shape facial recognition privacy on one dated timeline, each with its primary document.

Definition and terms defines face recognition and 25 working terms in one sentence each, from NIST FRTE pages, NISTIR 8280 and the ISO/IEC 19795 and 30107 standards.

Dataset withdrawals traces what happened after MS-Celeb-1M was taken down and MegaFace decommissioned, with the counts of copies, derivatives and models that stayed in use.

Face embeddings explains what a face embedding vector holds, how FaceNet and ArcFace train it, and how a distance or cosine threshold turns two vectors into a match decision.

NIST report cards walks through the header, the datasets, the operating points and the demographic panels of a NIST FRTE report card, and lists the reading mistakes that produce false rankings.