The VGGFace2 dataset is a large-scale face dataset containing 3.31 million images of 9131 subjects, with an average of 362.6 images per subject. The images were collected from Google Image Search and exhibit significant variations in pose, age, illumination, ethnicity, and profession. The dataset was designed to cover a wide range of pose, age, and ethnicity variations, and to minimize label noise. The dataset includes pose and age annotations, allowing for evaluation of face recognition across different poses and ages. The dataset was collected through a multi-stage process involving automated and manual filtering to ensure high accuracy. The dataset is divided into two splits: one for training (8631 classes) and one for evaluation (500 classes). The VGGFace2 dataset was used to train ResNet-50 and SE-ResNet-50 convolutional neural networks, which achieved state-of-the-art performance on the IJB benchmark datasets, surpassing previous results. The dataset and models are publicly available. The VGGFace2 dataset provides a comprehensive collection of face images with diverse variations, making it suitable for evaluating face recognition algorithms under various conditions. The dataset includes detailed annotations for pose and age, enabling the evaluation of face recognition performance across different poses and ages. The dataset was collected through a multi-stage process involving automated and manual filtering to ensure high accuracy. The dataset is divided into two splits: one for training (8631 classes) and one for evaluation (500 classes). The VGGFace2 dataset has been shown to significantly improve face recognition performance, particularly in handling pose and age variations. The dataset is a valuable resource for researchers working on face recognition and related tasks.The VGGFace2 dataset is a large-scale face dataset containing 3.31 million images of 9131 subjects, with an average of 362.6 images per subject. The images were collected from Google Image Search and exhibit significant variations in pose, age, illumination, ethnicity, and profession. The dataset was designed to cover a wide range of pose, age, and ethnicity variations, and to minimize label noise. The dataset includes pose and age annotations, allowing for evaluation of face recognition across different poses and ages. The dataset was collected through a multi-stage process involving automated and manual filtering to ensure high accuracy. The dataset is divided into two splits: one for training (8631 classes) and one for evaluation (500 classes). The VGGFace2 dataset was used to train ResNet-50 and SE-ResNet-50 convolutional neural networks, which achieved state-of-the-art performance on the IJB benchmark datasets, surpassing previous results. The dataset and models are publicly available. The VGGFace2 dataset provides a comprehensive collection of face images with diverse variations, making it suitable for evaluating face recognition algorithms under various conditions. The dataset includes detailed annotations for pose and age, enabling the evaluation of face recognition performance across different poses and ages. The dataset was collected through a multi-stage process involving automated and manual filtering to ensure high accuracy. The dataset is divided into two splits: one for training (8631 classes) and one for evaluation (500 classes). The VGGFace2 dataset has been shown to significantly improve face recognition performance, particularly in handling pose and age variations. The dataset is a valuable resource for researchers working on face recognition and related tasks.