TAR at FAR explained: one IJB-C row worked through, pair by pair

This page is part of the face recognition section and takes one row from the face verification metrics page apart. The same quantities carry NIST's names, FNMR at a fixed FMR, on the page about the four error rates.

01What do TAR and FAR measure?

A verification system compares two faces and returns a similarity score. Every pair in a test set is either genuine, two images of the same person, or impostor, two images of different people. Choose a threshold: genuine pairs scoring at or above it are accepted, which is the true accept rate (TAR); impostor pairs scoring at or above it are also accepted, which is the false accept rate (FAR). Raising the threshold lowers both rates together, so neither number means anything on its own. A published result fixes the FAR, finds the threshold that produces it, and reports the TAR at that threshold. The task itself is now defined in law as well as in papers: China's 2025 facial recognition measures define verification as a one-to-one comparison of a captured face with a stored one, as the page on face recognition rules in China records.

NIST states the same definition in its own vocabulary on the FRTE 1:1 page: "FNMR is the proportion of mated comparisons below a threshold set to achieve the false match rate (FMR) specified. FMR is the proportion of impostor comparisons at or above that threshold" (NIST FRTE 1:1). TAR is one minus FNMR; FAR is FMR; "mated" is genuine. The same FNMR appears in a full-system liveness test, which the page on liveness conformance letters describes. The academic papers and the NIST evaluations report the same curve from opposite ends.

02What is in the IJB-C verification row?

IJB-C's README (NIST, December 2017) defines Test 1, 1:1 verification, with four counts: 3,531 subjects, 23,124 templates, 19,557 genuine matches and 15,638,932 impostor matches (IJB-C README). A template is a set of images and video frames of one subject, pooled into one representation; the comparisons are between templates, not single images. The Partial FC paper describes the same protocol from the user's side: "it comprises 23,124 templates, 19,557 genuine pairs, and 15.6M impostor pairs", with each template built by weighting image and frame features by detection score and feature norm (arXiv 2010.05222).

The two pair counts are the whole story of the metric. There are about 800 impostor pairs for every genuine pair, which is why the FAR has to be small: at FAR 1e-2, which would sound strict, 156,389 impostor pairs would be accepted, eight times the number of genuine pairs in the set.

03The worked row: 97.2 percent TAR at FAR 1e-4

Take the row from the Partial FC paper's Table 2: ResNet-100 trained on Glint360K with CosFace loss and a Partial FC sampling rate of 0.1, IJB-C TAR 97.2 percent at FAR 1e-4 (the abstract's "97.2% TAR on IJB-C at FAR=10^-4"). Reading it pair by pair:

So the threshold is placed where 1,564 of the different-person pairs still look alike to the model, and at that threshold 548 same-person pairs look different. Move the threshold up until only 156 impostor pairs pass, FAR 1e-5, and more genuine pairs fall below it; the TAR drops. The number 97.2 describes the model only together with the number 1e-4.

04How much does the TAR move when the FAR moves?

The InsightFace ArcFace training README publishes IJB-C at both operating points for every model it trains, which makes the movement visible on the same model (arcface_torch README). ResNet-50 trained on Glint360K scores 97.16 at FAR 1e-4 and 95.81 at FAR 1e-5. ResNet-100 on MS1MV2 scores 96.37 and 94.27. The small MobileFaceNet on MS1MV2 scores 93.61 and 90.28, a drop of 3.3 points for the same tenfold tightening that costs the larger models 1.2 to 2.1 points. The figure above shows five rows.

Pushing further costs more. The RetinaFace paper reports an ArcFace model at 89.59 percent TAR on IJB-C at FAR 1e-6 (arXiv 1905.00641), while the ArcFace paper's own Table 7 gives 95.6 percent at FAR 1e-4 for its MS1MV2 ResNet-100 model (arXiv 1801.07698); those are different pipelines, so the gap is indicative rather than a single curve, but the direction is the same on every system. The ArcFace page carries the full Table 7.

05Why do two papers quote different TARs for IJB-C?

Four conditions change the number without changing the metric. The FAR: 1e-4 is the IJB-C convention in the papers on this site, while the ICCV 2021 masked face challenge reports TAR at FAR 1e-6, and the Partial FC paper's best MFR figure, 94.0 percent, is at that stricter point. The training set: in the Partial FC paper's Table 1 the same IR50 backbone scores 87.2 on IJB-C when trained on CASIA-WebFace and 97.1 when trained on WebFace12M. The template construction: feature weighting by detection score and norm, as both the ArcFace and Partial FC papers describe. And the evaluation code: the InsightFace README figures above come from the toolkit's own evaluation, the InsightFace page on this site describes it, and a README row and a paper row for a similar backbone and training set need not match, because loss, sampling rate and schedule can differ.

No single TAR predicts how a system performs at a different FAR, on a different test set, or in a deployment whose impostor pairs look nothing like IJB-C's.