VGGFace2 at a Glance
VGGFace2 was introduced by Qiong Cao, Li Shen, Weidi Xie, Omkar Parkhi and Andrew Zisserman in a paper at the 2018 IEEE Conference on Automatic Face and Gesture Recognition. The key numbers:
| VGGFace2 | |
|---|---|
| Images | 3.31 million |
| People | 9,131 |
| Images per person | 87 to 843, average about 362 |
| Training split | 8,631 people |
| Test split | 500 people |
| Source | Google Image Search |
| Extra labels | Pose and age templates for a subset of people |
The people are mostly public figures (actors, athletes, politicians) chosen to cover a wide range of professions, ages and ethnicities.
What Made It Different
By 2017, face datasets had two main shapes. Some, like CASIA-WebFace (about 500,000 images of 10,575 people), had many people but few photos each. Others, like the original VGGFace (2.6 million images of 2,622 people, 2015), had many photos each but fewer people.
VGGFace2 aimed for both: a decent number of people and a lot of photos per person, with deliberate variety. Its authors wanted every identity to appear across a wide range of:
- Pose, from front-on to near profile
- Age, with photos from different points in people's careers
- Lighting and image quality
- Ethnicity and profession across the identity list
That variety is what helps a model cope with real photos rather than neat portraits.
Read more: Why Training Data Is the Most Important Factor in Face Recognition Quality
How It Was Built
The team started from a list of names, downloaded images from Google Image Search for each, and then cleaned the results in stages. Automatic filtering used a face classifier to throw out images that didn't match the rest of a person's set. Human annotators then checked the remaining images. The paper reports that this kept label noise low, which was a big step up from some earlier web-scraped sets.
Near-duplicate images were removed, and the test set's 500 identities were kept completely separate from training, so evaluation reflected faces the model had never seen.
How It Compares With Other Datasets
| Dataset | Year | Images | People | Notes |
|---|---|---|---|---|
| LFW | 2007 | 13,233 | 5,749 | Evaluation only. Now too easy |
| CASIA-WebFace | 2014 | ~494,000 | 10,575 | Early large training set |
| VGGFace | 2015 | 2.6 million | 2,622 | Many photos per person, few people |
| MS-Celeb-1M | 2016 | ~10 million | 100,000 | Very noisy labels; withdrawn by Microsoft in 2019 |
| VGGFace2 | 2018 | 3.31 million | 9,131 | Deep and varied per person |
| MS1MV2 | 2019 | 5.8 million | 85,742 | Cleaned MS-Celeb-1M used by ArcFace; Ollie's training set |
The trade-off is width against depth. MS1MV2 has many more people, which tends to help models separate identities. VGGFace2 has more photos per person and more variety in each, which helps models handle pose and age.
Read more: Face Recognition Benchmarks Explained: LFW, IJB-C, and Beyond
Get new guides by email
Why Ollie Used a Wider Dataset Instead
For learning to tell people apart, the number of identities matters a lot. Losses like ArcFace and CosFace work by pushing each person's photos into their own region of embedding space, and seeing tens of thousands of people gives the model much more practice at that than 9,000. MS1MV2's 85,742 identities made it the standard training set for those losses, and the published results were easy to compare against.
So Ollie's 20-layer network was trained from scratch on a dataset with that many identities and reached 99.5% on LFW, with all LFW identities removed from the training data first. VGGFace2 remains a good choice when pose and age variation matter most, and many papers report results on both. More on why training data choices matter in why training data matters.
Licensing and Ethics
VGGFace2 was released for non-commercial research, and like most web-scraped face datasets, the people in it did not individually consent. Since around 2019, several large face datasets have been withdrawn or restricted after criticism over consent and use, and the official VGGFace2 download page has at times been unavailable. Anyone planning to use it should check its current licence terms and the rules where they work.
This is one reason Ollie's celebrity search index is built only from freely licensed Wikimedia Commons photos, with every photographer credited. See how Ollie's celebrity database was built.