The WebFace260M paper explained: data, cleaning, protocol and results

This page is the legacy strand of the face recognition section. The WebFace260M dataset page records how the data was distributed from this domain and the ICCV 2021 challenge page covers the competition built on it; this page reads the paper itself.

01What did the paper set out to do?

The paper's premise is that academic face recognition was data-limited. Its Table 1 lists the public training sets of the time and names MegaFace2 (4.7 million faces of 672,000 subjects) and MS1M as the largest by identities and by faces, while noting that companies train on far larger private collections. The authors' aim was a public set an order of magnitude larger, a cleaning method that needed no manual labour at that scale, and an evaluation that measured a whole recognition system under a fixed inference time rather than an embedding network alone.

Three contributions follow: the WebFace260M and WebFace42M data, the CAST pipeline, and the FRUITS protocol with a new attribute-labelled test set. A fourth, less often cited, is the distributed training work that made a ResNet-100 trainable on 42 million faces in a practical time.

02How were 260 million faces collected?

The identity list came from IMDB. The authors report that many listed names had no public images, so 3 million celebrity names were kept, and faces were downloaded through image search: 200 images per identity for the top 10 percent of subjects, then 100, 50 and 25 for the next 20, 30 and 40 percent. The raw download held 4,073,509 identities and 265,777,598 images (Table 2).

Faces were detected and aligned with five landmarks from RetinaFace, keeping only the largest face above a confidence threshold in each image. After that step the set held 4,008,130 identities and 260,890,076 faces, the WebFace260M of the title. The paper describes subjects from more than 200 countries or regions and more than 500 professions, with dates of birth back to 1846.

03How does CAST clean the data?

CAST stands for Cleaning Automatically utilizing Self-Training. The paper's argument is that MS1M was never cleaned and carried close to 50 percent noise, that the semi-automatic cleaning of VGGFace2 and IMDB-Face does not scale, and that clustering methods assume a clean set to begin with. The alternative is a teacher and student loop in three numbered steps: (1) a teacher model, a ResNet-100 trained with ArcFace on MS1MV2, cleans the original 260 million images; (2) a student of the same architecture is trained on the cleaned images and, having seen more data, embeds faces better than the teacher; (3) the student becomes the teacher and the loop repeats until the 42 million faces remain. Each pass cleans the original WebFace260M again with the new teacher.

Each pass has two parts. Intra-class cleaning works folder by folder: because image search returns one dominant subject per query, the 512-dimensional embeddings in each folder are clustered with DBSCAN and only the largest cluster of more than two faces is kept. Inter-class cleaning compares the feature centre of every folder with every other: two folders are merged when their cosine similarity is above 0.7, and the smaller folder is deleted when the similarity falls between 0.5 and 0.7. After three iterations, duplicates within a subject are removed above cosine 0.95, and any subject whose centre is within 0.7 of a face in LFW and its variants, FaceScrub, IJB-C or the paper's own test set is removed. The figure above follows the counts in Table 2 from 265.8 million faces to the final 2,059,906 identities and 42,474,558 faces.

04What is the FRUITS protocol?

FRUITS, Face Recognition Under Inference Time conStraint, tests the whole system rather than the embedding network: detection, alignment, embedding and matching must all finish within a time budget, measured on one core of an Intel Xeon E5-2630 v4 at 2.20 GHz. The paper names NIST's evaluation as the only time-constrained protocol of the time and notes that its submission cadence of one entry every four calendar months limits how freely academics can test. The face recognition test page sets this protocol beside the NIST and ISO conventions.

The metric is FNMR at a fixed FMR, defined in the paper as NIST defines it: the proportion of mated comparisons below a threshold set to give the specified false match rate. The test set was built by hand rather than drawn from public benchmarks, which the authors call saturated and noisy: 2,225 identities and 38,578 faces with age, race, gender and controlled-or-wild labels, giving 427,759 genuine and 743,683,994 impostor pairs on the All setting (Table 3). The paper says the set will be maintained and updated, which is the usual fate of a verification set once models have learned its faces.

05What do the results tables show?

Table 5 trains a ResNet-100 with three losses on each candidate training set and scores it on the same tests. On IJB-C at TAR at FAR 0.0001, the ArcFace model reaches 96.03 percent from MS1MV2, 96.77 from WebFace4M (10 percent of the cleaned data), 97.47 from WebFace12M (30 percent) and 97.70 from WebFace42M. The paper states the IJB-C gain as a relative 40 percent reduction in error rate, from 96.03 to 97.70. On the paper's own test set, FNMR at FMR 0.00001 falls from 5.08 percent for MS1MV2 to 1.58 percent for WebFace42M, which the paper describes as a relative improvement near 70 percent. The abstract adds that a model trained on WebFace42M "ranks the 3rd among 430 entries on NIST-FRVT" at the time; the abstract does not state which NIST measure produced the rank.

Two readings matter. First, WebFace4M, a tenth of the data, already beats every MS1M variant and MegaFace2 in the table, which is the paper's evidence that cleaning counts as much as scale. Second, every comparison holds the backbone, loss and test fixed, so the differences are attributable to training data alone; the same rule governs how this site reads detector tables, as the bridge page between the two sections explains.

06What happened to the benchmark after the paper?

The ICCV 2021 Masked Face Recognition Challenge ran a WebFace260M track on a new test set of 2,478 celebrities and 60,926 faces plus a real-world masked set; 69 teams submitted 833 solutions in its first phase and 49 beat the organisers' baseline (arXiv 2108.07189). The data was distributed from this domain under a licence agreement until the original site went offline. This site is not the WebFace260M project and is not affiliated with its authors, Tsinghua University or XForwardAI. The deep learning face recognition page places the paper among the training sets and losses that preceded it.