VGGFace2: A Dataset for Recognising Faces across Pose and Age

By Qiong Cao, Li Shen, Weidi Xie, Omkar M. Parkhi, Andrew Zisserman. 2017

3,096 citations recorded by the source corpus.

In this paper, we introduce a new large-scale face dataset named VGGFace2. The dataset contains 3.31 million images of 9131 subjects, with an average of 362.6 images for each subject. Images are downloaded from Google Image Search and have large variations in pose, age, illumination, ethnicity and profession (e.g. actors, athletes, politicians). The dataset was collected with three goals in mind: (i) to have both a large number of identities and also a large number of images for each identity; (ii) to cover a large range of pose, age and ethnicity; and (iii) to minimise the label noise. We describe how the dataset was collected, in particular the automated and manual filtering stages to ensure a high accuracy for the images of each identity. To assess face recognition performance using the new dataset, we train ResNet-50 (with and without Squeeze-and-Excitation blocks) Convolutional Neural Networks on VGGFace2, on MS-Celeb-1M, and on their union, and show that training on VGGFace2 leads to improved recognition performance over pose and age. Finally, using the models trained on these datasets, we demonstrate state-of-the-art performance on the IJB-A and IJB-B face recognition benchmarks, exceeding the previous state-of-the-art by a large margin. The dataset and models are publicly available.

Subfields: Computer Vision and Pattern Recognition, Evaluation & Benchmarks, Signal Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

In Evaluation & Benchmarks

In Signal Processing

Citations in the corpus

FairFace: Face Attribute Dataset for Balanced Race, Gender, and Age (2019, background): “To date, numerous large scale face image datasets [22, 31, 13, 69, 37, 24, 41, 68, 15, 27, 46, 8, 39] have been proposed and fostered research and development for automated face detection [35, 21], alignment [66, 44], recognition [57, 49], generation [67, 5, 26, 58], modiﬁcation [3, 32, 19], and…”
Make-A-Scene: Scene-Based Text-to-Image Generation with Human Priors (2022, methodology): “where the index l is used to denote the size of the spatial activation at specific layers of the face embedding network FE [6], while the summation runs over the last layers of each block of size 112× 112, 56× 56, 28× 28, 7× 7, 1× 1 (1 × 1 being the size of the top most block), ĉf and cf are respect”
This Face Does Not Exist... But It Might Be Yours! Identity Leakage in Generative Models (2020, methodology): “Model Name Architecture Training Data (# of images) Feature Vector Length Comparison Score face recognition [1] ResNet-34 Custom (∼3M) 128 Euclidean Distance FaceNet [23] GoogLeNet-24 VGGFace2 [4] (∼3.”
MERLOT RESERVE: Neural Script Knowledge through Vision and Language and Sound (2022, methodology): “[17] Qiong Cao, Li Shen, Weidi Xie, Omkar M Parkhi, and Andrew Zisserman.”
Mitigating Sybils in Federated Learning Poisoning (2018, methodology): “with minimal change to the original federated learning procedure. Moreover, FoolsGold does not assume a speciﬁc number of attackers. We evaluate FoolsGold on 4 diverse data sets (MNIST [24], VGGFace2 [13], KDDCup99 [16], Amazon Reviews [16]) and 3 model types (1-layer Softmax, Squeezenet, VGGNet) an”
Backdooring Convolutional Neural Networks via Targeted Weight Perturbations (2018, methodology): “• Experiments conducted over a five-layer CNN [16] and ResNet-50 [14] trained on the VGGFace2 dataset [2], with the latter assessment highlighting the attack’s effectiveness against face recognition networks.”
Model-Reuse Attacks on Deep Learning Systems (2018, methodology): “We then integrate the feature extractor of this model with a classifier (1 FC layer + 1 SM layer) and adapt the system to a dataset extracted from the VGGFace2 dataset [13], which consists of 25,000 facial images belonging to 500 individuals.”
Diversity in Faces (2019, methodology): “Several data sets such as IJBC [53], UMD [26] and VGGFace [22, 23] provide a larger set of face images with a wider range of pose and lighting variations.”
ScaIL: Classifier Weights Scaling for Class Incremental Learning (2020, background): “• VGGFace2 [4]: face recognition dataset including over 9000 identities in its full version.”
Eternal Sunshine of the Spotless Net: Selective Forgetting in Deep Networks (2019, methodology): “We report experiments on MNIST, CIFAR10 [18], Lacuna-10 and Lacuna-100, which we introduce and consist respectively of faces of 10 and 100 different celebrities from VGGFaces2 [6] (see Appendix for details).”
Compacting, Picking and Growing for Unforgetting Continual Learning (2019, methodology): “For face veriﬁcation, we use VGGFace2 [4] for training and LFW [18] for testing.”
IL2M: Class Incremental Learning With Dual Memory (2019, methodology): “VGGFace2 and Landmarks do not have standard test sets for IL.”
Class-Incremental Learning: Survey and Performance Evaluation on Image Classification (2020, methodology): “To study the effects of the different approaches on smaller domain shifts we use the VGGFace2 dataset [91].”
A Comprehensive Study of Class Incremental Learning Algorithms for Visual Tasks (2020, background): “5% while the corresponding drop for VGGFACE2 is only of 1 .”
Use What You Have: Video retrieval using representations from collaborative experts (2019, methodology): “Face embeddings are extracted in two stages: (1) Each frame is passed through an SSD face detector [4, 31] to extract bounding boxes; (2) The image region of each box is passed through a ResNet50 [22] that has been trained for the task of face classification on the VGGFace2 dataset [5].”
Turning a Blind Eye: Explicit Removal of Biases and Variation from Deep Neural Network Embeddings (2018, background): “Large publicly available face datasets are often composed of celebrities, such as [6,7,8,9,10, 11].”
SphereFace Revived: Unifying Hyperspherical Face Recognition (2021, methodology): “Our experiments show that that SFN generally yields comparable or even better results than HFN with a properly chosen t when our models are trained on VGGFace2.”
DeepFakes and Beyond: A Survey of Face Manipulation and Fake Detection (2020, background): “Researchers focused on this type of manipulation usually consider real faces from popular public databases such as CelebA [23], FFHQ [19], CASIA-WebFace [24], and VGGFace2 [25], among others to train their systems.”
Racial Faces in the Wild: Reducing Racial Bias by Information Maximization Adaptation Network (2018, methodology): “Then, we also examine some SOTA algorithms, i.e. Center-loss [65], Sphereface [39], VGGFace2 [15] and Ar-cFace [21], as well as four commercial recognition APIs, i.e. Face++, Baidu, Amazon, Microsoft on our RFW.”