Targeting Ultimate Accuracy: Face Recognition via Deep Embedding

By Jingtuo Liu, Yafeng Deng, Tao Bai, Chang Huang. 2015

245 citations recorded by the source corpus.

Face Recognition has been studied for many decades. As opposed to traditional hand-crafted features such as LBP and HOG, much more sophisticated features can be learned automatically by deep learning methods in a data-driven way. In this paper, we propose a two-stage approach that combines a multi-patch deep CNN and deep metric learning, which extracts low dimensional but very discriminative features for face verification and recognition. Experiments show that this method outperforms other state-of-the-art methods on LFW dataset, achieving 99.77% pair-wise verification accuracy and significantly better accuracy under other two more practical protocols. This paper also discusses the importance of data size and the number of patches, showing a clear path to practical high-performance face recognition systems in real world.

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

Diversity in Faces (2019, methodology): “8% accuracy [18,19].”
A Discriminative Feature Learning Approach for Deep Face Recognition (2016, background): “The effectiveness of triplet loss has been demonstrated in [21,25,27].”
Latent Factor Guided Convolutional Neural Networks for Age-Invariant Face Recognition (2016, background): “nowadays, deep convolution neural networks (CNNs) have been successfully applied to a variety of problems in computer vision, including object detection and classification [14, 23, 28], and face recognition [33, 32, 29, 20, 36], etc.”
ArcFace: Additive Angular Margin Loss for Deep Face Recognition (2018, unknown): “Method #Image LFW YTF DeepID [30] 0.2M 99.47 93.20 Deep Face [31] 4.4M 97.35 91.4 VGG Face [22] 2.6M 98.95 97.30 FaceNet [27] 200M 99.63 95.10 Baidu [15] 1.3M 99.13 - Center Loss [36] 0.7M 99.28 94.9 Range Loss [43] 5M 99.52 93.70 Marginal Loss [8] 3.8M 99.48 95.98 SphereFace [17] 0.5M 99.42 95.0…”
Striking the Right Balance With Uncertainty (2019, unknown): “DeepFace [48] 97.35 91.4 FaceNet [42] 99.63 95.4 Web-scale [49] 98.37 - DeepID2+ [47] 99.47 93.2 Baidu [29] 99.13 - Center Face [55] 99.28 94.9 Marginal Loss [13] 99.48 96.0 Noisy Softmax [8] 99.18 94.9”
Improved Deep Metric Learning with Multi-class N-pair Loss Objective (2016, background): “As to experimental results, only a few has reported strong empirical performance using these loss functions alone [19, 26], but many have combined with softmax loss to train deep networks [22, 31, 18, 14, 32].”
Channel-Level Acceleration of Deep Face Representations (2015, background): “Another contribution [11] suggests first training a feature extraction network followed by a metric learning network to reduce the feature to a low dimension.”
Deep Face Recognition: A Survey (2018, methodology): “Other methods combine triplet loss with softmax loss [239], [105], [46], [36].”
Trunk-Branch Ensemble Convolutional Neural Networks for Video-Based Face Recognition (2016, background): “hand, although pose variations and occlusion are partially solved in SIFR by ensemble modelling [6], [10], the strategy may not be directly extended to VFR.”
Deep Imbalanced Learning for Face Recognition and Attribute Prediction (2018, background): “Another popular choice is the triplet loss [33], [46], which leads to state-ofthe-art performance.”
When Face Recognition Meets with Deep Learning: An Evaluation of Convolutional Neural Networks for Face Recognition (2015, background): “Apart from choosing regions from four corners and the center, semantic regions are chosen for training networks by [17] and [24].”
Deep Metric Learning: A Survey (2019, background): “While decreasing the distance between dissimilar samples [27,28], deep metric learning, which aims to increase the distance between similar samples, is directly related to the distance between samples.”
L2-constrained Softmax Loss for Discriminative Face Verification (2017, methodology): “The proposed method yields a consistent and significant boost on all the three challenging face veriﬁcation datasets namely LFW [14], YouTube Face [19] and IJB-A [16] Moreover, the gains from L 2 -softmax loss are complementary to metric learning (eg: TPE [27], joint-Bayes [3]) or auxiliary loss…”
Past, Present, and Future of Face Recognition: A Review (2020, background): “[122] (2015) have proposed a two-stage approach for face recognition, which extracts low dimensional yet high discriminative features by merging a multi-patch deep CNN with deep metric learning.”
A Comprehensive Study on Center Loss for Deep Face Recognition (2019, unknown): “(2015);Liu et al. (2015).”