Face recognition datasets compared: identities, images and status

This page belongs to the face recognition section and is the data companion to the five benchmarks side by side page, which covers protocols and metrics. Here the question is the data itself: how many identities and images each set holds, what it is used for, and whether a new researcher can still obtain it.

01Which sets train a model and which test one?

Face recognition data plays three roles. A training set teaches a network to embed faces, so it needs many identities and many images per identity. A verification test set holds pairs of images labelled same or different, as LFW does. An identification test set holds a gallery and probes, so that a search returns a ranked list, as MegaFace and IJB-C do. A single dataset can serve two roles, but a published score is only meaningful when the training set and the test set share no identities.

That separation is why the authors of CASIA-WebFace removed the 1,043 subjects whose names matched LFW, as the CASIA-WebFace page records, and why the IJB-C readme says it is "designed to have no overlap" with CASIA WebFace and VGG-Face. The comparison below therefore labels each set by role.

02How do eight face recognition datasets compare?

The table orders the sets by year of publication. Counts are those the owners or the InsightFace README state; where a set has two counts, both are given.

The identity counts for MS-Celeb-1M need a note. Its paper proposes a task of recognising one million celebrities and its training set covers the 100,000 most frequent of them, which Exposing.ai records as about 100,000 people. The paper's abstract gives the image count as "10M images in version 1".

03Which of the eight can a new researcher still download?

Three owners have said so in writing. The MegaFace project page states that "MegaFace is being decommissioned and MegaFace data are no longer being distributed". NIST's face challenges page says it "discontinued distribution of all three IJB challenges (IJB-A, IJB-B, and IJB-C) on March 14th 2023". The VGGFace2 page at the University of Oxford says "The download links for the VGGFace2 dataset are no longer available from this website", although its 2018 paper had stated that the dataset and models were publicly available.

MS-Celeb-1M is a different case. Microsoft's site stopped serving it in 2019, but a 2021 study of citing papers found the data still on Academic Torrents and in derived datasets, which the withdrawn recognition sets page and the MegaFace page treat as part of one pattern. The face datasets and consent page gathers the stated reasons for each withdrawal. A withdrawn set can still be used by groups that hold a copy, so its numbers keep appearing in papers that nobody else can reproduce.

WebFace260M is the set this domain once hosted. Its paper describes the noisy collection and the cleaned WebFace42M, and the data was distributed from the original static site under a signed agreement until that site closed in January 2025. This site does not host it; the WebFace260M paper explained page reads the paper itself.

04How fast did training sets grow?

CASIA-WebFace had 494,414 images in 2014. WebFace42M, the cleaned set of 2021, has 42 million faces, about 85 times as many, and the noisy WebFace260M collection has 260 million, more than 500 times as many. In between, MS-Celeb-1M reached 10 million images in 2016 and Glint360K 17 million. The growth was in identities as well as images: CASIA-WebFace had 10,575 identities, Glint360K 360,000 and the cleaned WebFace42M 2 million.

Each jump came from a larger crawl of public photographs of known people and a cleaning step to reduce wrong labels. MS-Celeb-1M left its noise in; the cleaned MS1M-ArcFace version in the InsightFace README holds 85,000 identities and 5.8 million images, against about 100,000 people and 10 million images in the original, so cleaning removed about 42 percent of the images. The size of a set after cleaning is the figure to compare, not the size of the crawl.

05Does a larger training set give a better score?

On the Partial FC paper's controlled test, yes, with diminishing returns. The paper trains the same IR50 backbone on different sets and scores each on IJB-C as a true accept rate at a false accept rate of 10^-4: CASIA 87.2 percent, VGGFace 91.2, MS1MV2 96.1, MS1MV3 96.6, Glint360K 97.1 and WebFace12M 97.1 (arXiv 2010.05222). The TAR at FAR page defines the metric. The same table scores each model on the All subset of the ICCV21-MFR masked face challenge test at a false accept rate of 10^-6, where the spread is wider: 36.8 percent for CASIA, 77.7 for MS1MV2 and 90.6 for WebFace12M.

The lesson is about the test as much as the data. A harder protocol, a lower false accept rate or masked faces, separates training sets that an easy protocol cannot. On LFW, published accuracies span only 0.68 points from DeepID2 in 2014 to ArcFace in 2019, so the benchmark no longer separates systems, which is why the LFW dataset page treats it as a check that a model is not broken.

06What to record when citing a dataset

1. The exact set and version: MS1M-ArcFace and MS1MV3 are cleaned derivatives of MS-Celeb-1M, not the original.
2. The role: training, verification test or identification test.
3. The counts after cleaning, with the identities and the images given separately.
4. The date the owner's page or paper was read, because availability changes: four of the eight sets above are no longer distributed by their owners.
5. Whether the training and test sets share identities, and how that was checked.