Glint360K and MS1M: the training sets behind ArcFace models

This page belongs to the face recognition section and sits beside the VGGFace2 record, another training set in the same InsightFace list. Here the subject is the two derived sets that most open ArcFace-style models are trained on, and why their numbers matter when a result is quoted.

01What are Glint360K and MS1M?

MS1M is shorthand for versions of MS-Celeb-1M, a 2016 set of web images of public figures. The InsightFace dataset list carries several: MS1M-IBUG at 85K identities and 3.8 million images, and MS1M-ArcFace at 85K identities and 5.8 million, with references to the MS-Celeb-1M paper and to the marginal loss and ArcFace papers (InsightFace dataset README). The ArcFace paper names its version MS1MV2 and tabulates it at 85K identities and 5.8 million images (arXiv 1801.07698).

Glint360K is a later and larger set. The same list gives it 360K identities and 17 million images and credits the Partial FC paper, whose authors describe it as built by cleaning Celeb-500K and merging it with MS1MV2 (arXiv 2010.05222). A smaller sibling, Glint-Mini, is listed at 91K identities and 5.2 million images.

02How was Glint360K assembled?

The Partial FC paper states the recipe in one sentence: "We also clean Celeb-500K and merge it with MS1MV2 to construct Glint360K, containing 17 million images from 360K identities". Celeb-500K is a separate web collection that the Partial FC paper describes as containing roughly half a million identities and the InsightFace list gives as 500K identities and 50 million images.

That makes Glint360K partly a descendant of MS-Celeb-1M, because MS1MV2 is one of its two components. The point matters for the withdrawal question below.

03What do models trained on each score?

Partial FC's Table 1 trains one IR50 backbone on each set under the same schedule, and reports TAR on IJB-C at a false accept rate of 10^-4 and on the ICCV21-MFR test at 10^-6.

IJB-C saturates quickly: moving from MS1MV2 to Glint360K gains one point, while the harder ICCV21-MFR column gains 9.1. Table 2, with an IR100 backbone, lists Glint360K trained with CosFace and Partial FC at 96.1 on IJB-B and 97.2 on IJB-C, and MS1MV2 trained with ArcFace and Partial FC at 94.8 and 96.2; the loss differs between those two rows. The abstract's headline is 97.2 percent TAR on IJB-C at a false accept rate of 10^-4. The TAR at FAR page explains the metric, and the ArcFace page explains the loss that most of these rows use.

04Why do the counts differ between sources?

Three differences are worth knowing. The InsightFace list and the Partial FC paper agree on 360K identities and 17 million images for Glint360K, but the paper's current abstract page no longer carries the counts, and they appear only in its experiments section. The two MS1M versions differ by two million images, 3.8 million for MS1M-IBUG and 5.8 million for MS1M-ArcFace, because they were cleaned by different procedures. And the name MS1MV3, which Partial FC lists, has no count in the InsightFace list, so this page gives none.

A table that mixes them is therefore unsafe. A row labelled MS1M can mean either version, and a result of 96.1 on IJB-C is a statement about the set as much as about the network.

05Can these datasets still be used?

The code in InsightFace is released under an open licence, but its README says the training data containing the annotation, and the models trained with these data, "are available for non-commercial research purposes only". MS-Celeb-1M itself was withdrawn in April 2019, and an audit of citing papers found the original and five derived datasets still available in 2021 (Peng et al., arXiv 2108.02922). The withdrawn sets page covers what remained and the rights questions raised.

Because Glint360K includes MS1MV2, which is a cleaned subset of MS-Celeb-1M, the 2019 withdrawal is relevant to it as well as to the MS1M versions.

06How should a result trained on one of these sets be read?

Name the training set, then the backbone, then the loss. A row such as "IR100, ArcFace, MS1MV2" is not comparable with "IR100, CosFace, Glint360K" even when both use the same test set, because the data changed at the same time as the loss. The InsightFace toolkit page lists the other sets in the zoo, and the WebFace260M paper explained covers the larger 42 million image set that the paper built.