WebFace260M: A Benchmark for Million-Scale Deep Face Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Biometrics

Background:

• Face recognition systems require large-scale, high-quality datasets for effective training and evaluation.
• Existing benchmarks may not adequately represent the diversity and scale needed for state-of-the-art performance.
• Bridging the data gap between academic research and industrial applications is crucial.

Purpose of the Study:

• To introduce a new million-scale face recognition benchmark, WebFace260M (uncrurated) and WebFace42M (cleaned).
• To develop a scalable data cleaning pipeline (CAST) for massive face datasets.
• To establish a time-constrained evaluation protocol (FRUITS) and a masked face subset for comprehensive biometrics assessment.

Main Methods:

• Collected 4M name lists and downloaded 260M faces from the internet for WebFace260M.
• Developed the Cleaning Automatically utilizing Self-Training (CAST) pipeline to purify WebFace260M into WebFace42M.
• Constructed the Face Recognition Under Inference Time conStraint (FRUITS) protocol and a new test set with rich attributes.

Main Results:

• WebFace42M is the largest public face recognition training set, containing 2M identities and 42M faces.
• Reduced failure rate by 40% on the IJB-C set and achieved 3rd rank on NIST-FRVT using WebFace42M.
• Even a subset (WebFace4M) demonstrated superior performance over existing public training sets.

Conclusions:

• The proposed benchmark significantly advances face recognition research by providing unprecedented data scale and a robust evaluation framework.
• The benchmark demonstrates potential for standard, masked, and unbiased face recognition scenarios.
• The findings highlight the importance of large-scale, clean datasets for achieving state-of-the-art performance in face recognition.