VGGFace2 Large-Scale Face Image Collection

VGGFace2 is a large-scale dataset of facial images for recognition research. It contains over 3.1 million images of more than 9,000 subjects, created by the Visual Geometry Group at the University of Oxford. The dataset was released to advance face recognition model training.

Use Cases

• Train face recognition models using the subject identity labels and cropped-aligned image files.
• Benchmark face verification performance on the provided test sets with matched and mismatched identity pairs.
• Analyze pose and age attributes using the provided metadata annotations for each subject.
• Study dataset bias and fairness by examining the distribution of identities across age, ethnicity, and pose.

Strengths

• Over 3.1 million facial images providing substantial training data.
• Covers more than 9,000 distinct identities for diverse recognition tasks.
• Images include large variations in pose, age, lighting, and ethnicity.

Limitations

• Images were collected from the internet, potentially introducing label noise and selection bias.
• The dataset may have imbalanced representation across demographic attributes like ethnicity and age.
• Some images are of low resolution or contain occlusions, complicating model training.

Provenance

Source

Visual Geometry Group, University of Oxford.

Collection Method

Images crawled from the internet and annotated with identity and attributes.

Time Range

null

Freshness

null

Geography

Global, based on internet sources.