WebFace260M is a million-level face benchmark. The dataset consists of 4M identities and 260M faces, providing a good resource for deep face cleaning and recognition at the million level. The benchmark is designed for the research community to narrow the data gap behind the industry.

The dataset consists of the following data:

• Data with lots of noise: 4 million identities and 260 million faces;
• High-quality training data obtained through automatic cleaning: 42 million images involving 2 million identities;
• Property-rich test sets and time-constrained evaluation protocols.