VGGFace2: 3.31 million images of 9,131 subjects, built for pose and age

This page belongs to the face recognition section and records one of the training sets that open models draw on. Its larger contemporary, MS-Celeb-1M, had about 10 million images of 100,000 people; VGGFace2 chose depth per identity over breadth of identities, and the later Glint360K and MS1M sets went the other way.

01What does VGGFace2 contain?

The paper's abstract describes images "downloaded from Google Image Search" with "large variations in pose, age, illumination, ethnicity and profession (e.g. actors, athletes, politicians)" (arXiv 1710.08092). Each subject has between 80 and 843 images, with 362.6 on average, and the set is described as approximately gender-balanced, with 59.3 percent males. The authors say it includes more Chinese and Indian faces than the earlier VGGFace, while noting that "the ethnic balance is still limited by the distribution of celebrities and public figures".

The release includes human-verified bounding boxes, five landmark points per face, and estimated pose and apparent age. Its stated design goals were a large number of identities, a large number of images per identity, wide coverage of pose, age and ethnicity, and minimal label noise.

02How was it collected and filtered?

The pipeline had six stages. An initial list of 500,000 public figures came from the Freebase knowledge graph, and annotators kept names for which about 90 percent or more of 100 test images showed the right person, leaving 9,244 subjects. For each, 1,000 images were downloaded from Google Image Search, plus 200 for the added keyword "sideview" and 200 for "very young", to pull in profile views and younger ages: 1,400 per name, 12.94 million in all.

Face detection left 7.31 million images, a classifier-based filter 6.99 million and near-duplicate removal 5.45 million. A final automatic and manual pass, which took 21 days of annotation, left 3.31 million images. In that stage the authors removed 19 classes that overlapped with others and 94 subjects with fewer than 80 images, which gives the final 9,131 subjects. They aimed for a label purity above 96 percent.

03What pose and age coverage does it state?

Pose and age were not hand labelled for every image. The authors trained a five-way ResNet-50 on CASIA-WebFace to predict head pose and an eight-way ResNet-50 on IMDB-WIKI to predict apparent age, then applied both to the whole set. The pose and age histograms in the paper are therefore estimates from those classifiers.

For evaluation, the test split carries template annotations. A pose template is five faces of one subject with a consistent pose, frontal, three-quarter or profile; 300 subjects have two templates per view, giving 1.8K templates and 9K images. An age template is five faces with an apparent age below 34, deemed young, or 34 and above, deemed mature; 100 subjects have two templates per period, giving 400 templates and 2K images. Matching across young and mature faces was found substantially harder than within either group.

04How did models trained on it score?

The paper trained ResNet-50 and SE-ResNet-50 networks on VGGFace2, on MS-Celeb-1M and on both. On IJB-B at a false accept rate of 10^-4, a ResNet-50 trained on VGGFace2 reached a true accept rate of 0.784, against 0.743 for MS-Celeb-1M and 0.535 for the earlier VGGFace set.

On IJB-C at the same operating point the VGGFace2 ResNet-50 scored 0.825, the SE-ResNet-50 0.840 and the fine-tuned SE-ResNet-50 0.862. The paper reads the gain over MS-Celeb-1M as evidence of data quality, noting that MS-Celeb-1M has "tenfold more subjects and threefold more images". The TAR at FAR page explains the metric. These are 2018 results; the ArcFace page shows models trained on larger sets scoring higher on the same IJB-C protocol.

05Why do the counts differ between sources?

The paper says 3.31 million images and 9131 subjects. The InsightFace dataset list gives "9K ids/3.31M images", and the ArcFace paper's Table 1 rounds to 9.1K identities and 3.3M images. These are the same count at different precision. A different figure to avoid is 9,244 subjects, which is the pipeline's count before the last stage removed 113.

06Is VGGFace2 still available?

A 2021 audit lists VGGFace2 among the research datasets whose creators retracted them: "These concerns have led to the retractions of prominent research datasets including Tiny Images, VGGFace2, DukeMTMC, and MS-Celeb-1M" (Peng et al., arXiv 2108.02922). The InsightFace toolkit still lists it among training sets and states that data and models trained on it are for non-commercial research only. The withdrawn sets page covers the audit.