The paper introduces VGGFace2, a large-scale face dataset designed to address the challenges of recognizing faces across pose and age variations. The dataset contains 3.31 million images of 9131 subjects, with an average of 362.6 images per subject. Images were collected from Google Image Search and cover a wide range of pose, age, illumination, ethnicity, and profession. The dataset was collected with three main goals: to have a large number of identities and images per identity, to cover a broad range of pose, age, and ethnicity, and to minimize label noise. The authors describe the dataset collection process, which includes automated and manual filtering stages to ensure high accuracy. To evaluate the performance of face recognition using VGGFace2, the authors trained ResNet-50 and SE-ResNet-50 (SENet) on VGGFace2, MS-Celeb-1M, and their union. They found that training on VGGFace2 significantly improved recognition performance over pose and age. The models trained on these datasets demonstrated state-of-the-art performance on the IJB datasets (IJB-A, IJB-B, and IJB-C), outperforming previous methods by a large margin. The dataset and models are publicly available for research purposes.The paper introduces VGGFace2, a large-scale face dataset designed to address the challenges of recognizing faces across pose and age variations. The dataset contains 3.31 million images of 9131 subjects, with an average of 362.6 images per subject. Images were collected from Google Image Search and cover a wide range of pose, age, illumination, ethnicity, and profession. The dataset was collected with three main goals: to have a large number of identities and images per identity, to cover a broad range of pose, age, and ethnicity, and to minimize label noise. The authors describe the dataset collection process, which includes automated and manual filtering stages to ensure high accuracy. To evaluate the performance of face recognition using VGGFace2, the authors trained ResNet-50 and SE-ResNet-50 (SENet) on VGGFace2, MS-Celeb-1M, and their union. They found that training on VGGFace2 significantly improved recognition performance over pose and age. The models trained on these datasets demonstrated state-of-the-art performance on the IJB datasets (IJB-A, IJB-B, and IJB-C), outperforming previous methods by a large margin. The dataset and models are publicly available for research purposes.