VGGFace2: A Dataset for Recognising Faces across Pose and Age

Abstract

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 VG-GFace2, 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 face recognition of IJB datasets, exceeding the previous state-of-the-art by a large margin. The dataset and models are publicly available 1 .

Compact Deep Aggregation for Set Retrieval

“…Note, in all the experiments, there is no overlap between the query identities used for testing and the identities used for training the network, as the VGG Face Dataset (used for testing, e.g. for forming the Celebrity Together dataset) and the VGGFace2 Dataset (Cao et al, 2018) (used for training) share no common identities. Evaluation protocol.…”

“…Training data. The network is trained using faces from the training partition of the VGGFace2 dataset (Cao et al, 2018). This consists of 8631 identities, with on average 360 face samples for each identity.…”

“…A good (and necessary) initialization for the network is obtained as follows. The face feature extraction block is pretrained for single face classification on the VGGFace2 Dataset (Cao et al, 2018) using softmax loss. The NetVLAD layer, with K = 8 clusters, is initialized using k-means as in .…”

“…The exceptions are Baseline and Baseline+W, which simply use the feature extractor network with averagepooling, but no training. It is important to note that training the modified ResNet-50 and SENet-50 networks on the VGGFace2 dataset (Cao et al, 2018) provides very strong baselines (referred to as Baseline). Although the objective of this paper is not face classification, to demonstrate the performance of this baseline network, we test it on the public IARPA Janus Benchmark A (IJB-A dataset) (Klare et al, 2015).…”

“…Although the objective of this paper is not face classification, to demonstrate the performance of this baseline network, we test it on the public IARPA Janus Benchmark A (IJB-A dataset) (Klare et al, 2015). Specifically, we follow the same test procedure for the 1:N Identification task described in (Cao et al, 2018). In terms of the true positive identification rate (TPIR), the ResNet-50 128-D network achieves 0.975, 0.993 and 0.994 for the top 1, 5 and 10 ranking respectively, which is on par with the state-of-the-art networks in (Cao et al, 2018).…”