Web-scale training for face identification

By Yaniv Taigman, Ming Yang, Marc'Aurelio Ranzato, Lior Wolf. 2014

269 citations recorded by the source corpus.

Scaling machine learning methods to very large datasets has attracted considerable attention in recent years, thanks to easy access to ubiquitous sensing and data from the web. We study face recognition and show that three distinct properties have surprising effects on the transferability of deep convolutional networks (CNN): (1) The bottleneck of the network serves as an important transfer learning regularizer, and (2) in contrast to the common wisdom, performance saturation may exist in CNN's (as the number of training samples grows); we propose a solution for alleviating this by replacing the naive random subsampling of the training set with a bootstrapping process. Moreover, (3) we find a link between the representation norm and the ability to discriminate in a target domain, which sheds lights on how such networks represent faces. Based on these discoveries, we are able to improve face recognition accuracy on the widely used LFW benchmark, both in the verification (1:1) and identification (1:N) protocols, and directly compare, for the first time, with the state of the art Commercially-Off-The-Shelf system and show a sizable leap in performance.

Subfields: Computer Vision and Pattern Recognition, 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 Signal Processing

Citations in the corpus

How Image Degradations Affect Deep CNN-Based Face Recognition? (2016, methodology): “For example, the rank-1 face identification result with the single model version of the approach presented in [18] is reported as %72.”
Low-Shot Learning from Imaginary Data (2018, methodology): “Con-trastive loss functions [12, 16] and variants of the triplet loss [31, 29, 8] have been used for learning feature representations suitable for low-shot learning; the idea is to push examples from the same class closer together, and farther from other classes.”
TADAM: Task dependent adaptive metric for improved few-shot learning (2018, methodology): “In the same vein, in [5, 26, 30], triplet loss is used for learning a representation for few-shot learning.”
VGGFace2: A Dataset for Recognising Faces across Pose and Age (2017, methodology): “For instance, Facebook [21] trained a face identification model using 500 million images of over 10 million subjects.”
Single Image 3D Interpreter Network (2016, background): “Some previous works also use this idea and achieve better performance with lower computational cost in object detection [40] and face recognition [42].”
The MegaFace Benchmark: 1 Million Faces for Recognition at Scale (2015, methodology): “Other recent studies have considered both identification as well as verification results on LFW [2, 30, 28, 27].”
MS-Celeb-1M: A Dataset and Benchmark for Large-Scale Face Recognition (2016, methodology): “The publicly available datasets are much smaller than that being used privately in industry, such as Facebook [2,3] and Google [4], as summarized in Table 1.”
AdvHat: Real-World Adversarial Attack on ArcFace Face ID System (2019, background): “Last years face recognition systems based on deep learning and massive training data provided a very high level of recognition which outperforms human level of veriﬁcation [1] as well as identiﬁcation [2].”
MegaFace: A Million Faces for Recognition at Scale (2015, methodology): “Recently, several papers reported identification results on the LFW dataset [3, 28, 26, 25].”
Low-Shot Visual Recognition by Shrinking and Hallucinating Features (2016, methodology): “[41] find that low-dimensional feature representations transfer better on faces and Galanti et al.”
Do We Really Need to Collect Millions of Faces for Effective Face Recognition? (2016, result): “Finally, a comparison of our results on LFW to those reported by methods trained on millions of images ([35,36,24] and [29]), shows that with the initial”
UMDFaces: An annotated face dataset for training deep networks (2016, methodology): “Recently, Facebook used a dataset of about 500 million images over 10 million identities for face identification [37].”
One-shot Face Recognition by Promoting Underrepresented Classes (2017, background): “Recently, we have seen a considerable amount of research results reported on LFW, which are obtained by using different algorithms on different datasets [1, 17, 16, 18, 15, 19, 20].”
Variational Few-Shot Learning (2019, methodology): “In addition, contrastive loss [17, 25] and triplet loss [40, 44] are used for strengthen learned metrics by fully exploring pair/triplet relationships within the dataset.”