SphereFace: Deep Hypersphere Embedding for Face Recognition

By Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, Le Song. 2017

3,098 citations recorded by the source corpus.

This paper addresses deep face recognition (FR) problem under open-set protocol, where ideal face features are expected to have smaller maximal intra-class distance than minimal inter-class distance under a suitably chosen metric space. However, few existing algorithms can effectively achieve this criterion. To this end, we propose the angular softmax (A-Softmax) loss that enables convolutional neural networks (CNNs) to learn angularly discriminative features. Geometrically, A-Softmax loss can be viewed as imposing discriminative constraints on a hypersphere manifold, which intrinsically matches the prior that faces also lie on a manifold. Moreover, the size of angular margin can be quantitatively adjusted by a parameter m. We further derive specific m to approximate the ideal feature criterion. Extensive analysis and experiments on Labeled Face in the Wild (LFW), Youtube Faces (YTF) and MegaFace Challenge 1 show the superiority of A-Softmax loss in FR tasks.

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

Unsupervised Feature Learning via Non-parametric Instance Discrimination (2018, methodology): “An important technical point on metric learning for face recognition is normalization [35, 22, 43], which we also utilize in this work.”
This Face Does Not Exist... But It Might Be Yours! Identity Leakage in Generative Models (2020, neutral): “3M) 512 Euclidean Distance ArcFace [7] ResNet-100 MS-Celeb-1M [12] (∼10M) 512 Cosine Distance SphereFace [19] ResNet-64 CASIA-Webface [30] (∼0.”
NTU RGB+D 120: A Large-Scale Benchmark for 3D Human Activity Understanding (2019, methodology): “(3), and then perform classiﬁcation using the nearest neighbour classiﬁer, as in [101].”
Continual Learning by Modeling Intra-Class Variation (2022, methodology): “To enable an effective modeling of ∆f , we draw inspirations from the observations in (Liu et al., 2017b; 2018b;a; Chen et al., 2020) that geodesic distance on hypersphere is well aligned with perceptual difference, and propose to model the distribution ∆f on the hypersphere in MOCA.”
Few-Shot Class-Incremental Learning from an Open-Set Perspective (2022, positive): “Inspired by the relation between normalized weights on the last fully connected layer and class centers, Liu et al. proposed SphereFace [14] which uses an angular margin penalty to enforce extra intra-class compactness and inter-class discrepancy.”
Few-Shot Class-Incremental Learning (2020, background): “Convolutional Neural Networks (CNNs) have been successfully applied to a broad range of computer vision tasks [17, 11, 34, 4, 22, 25, 45, 19].”
Maintaining Discrimination and Fairness in Class Incremental Learning (2019, methodology): “This phenomenon is mainly caused by class imbalance [9, 21].”
Compacting, Picking and Growing for Unforgetting Continual Learning (2019, methodology): “We use the 20-layer CNN in SphereFace [21] and train a model for the face verification task accordingly.”
Exploring the Encoding Layer and Loss Function in End-to-End Speaker and Language Recognition System (2018, background): “This procedure is non-trivial, both time-consuming and performance-sensitive [33].”
TVQA+: Spatio-Temporal Grounding for Video Question Answering (2019, background): “, 2016) and recognition model (Liu et al., 2017), but they do not work well mainly due to many partial occlusion of faces (e.”
Bi-Objective Continual Learning: Learning 'New' While Consolidating 'Known' (2020, background): “…(Krizhevsky, Sutskever, and Hinton 2012; He et al. 2015; Wang et al. 2010; Ren et al. 2015; Redmon et al. 2015; Cheng et al. 2016; Shun et al. 2016; Long, Shelhamer, and Darrell 2015; Chen et al. 2017; Liu et al. 2017; Deng et al. 2018; Tusng-Yu, Aruni, and Subhransu 2015; Wei et al. 2018a; 2018b).”
Understanding Contrastive Representation Learning through Alignment and Uniformity on the Hypersphere (2020, background): “Existing works have motivated this choice from a manifold mapping perspective (Liu et al., 2017; Davidson et al., 2018) and computation stability (Xu & Durrett, 2018; Wang et al., 2017).”
CosFace: Large Margin Cosine Loss for Deep Face Recognition (2018, methodology): “A-Softmax improves the softmax by introducing extra margin, such that the decision boundary is given by: C1 : cos(mθ1) ≥ cos(θ2) C2 : cos(mθ2) ≥ cos(θ1) Thus, for C1 it requires θ1 ≤ θ2m , and similarly for C2.”
Discrimination-aware Channel Pruning for Deep Neural Networks (2018, methodology): “Furthermore, pruning 65% channels in SphereNet-4 results in a more compact model, which requires only 24.16M FLOPs with the accuracy of 98.02% on LFW.”
Expression, Affect, Action Unit Recognition: Aff-Wild2, Multi-Task Learning and ArcFace (2019, methodology): “This, as well as the recent tendency to develop elaborate loss functions for speciﬁc tasks [18], has led us to search In fact, in the related face recognition ﬁeld, it has been shown [24, 44] that categorical cross entropy loss is insufﬁcient to acquire discriminating power for face classiﬁcation.”
Larger Norm More Transferable: An Adaptive Feature Norm Approach for Unsupervised Domain Adaptation (2018, methodology): “This technique is widely used to characterize the feature embeddings under the softmax-related objectives [47, 21, 51].”
Mitigating Neural Network Overconfidence with Logit Normalization (2022, methodology): “…the literature, normalization has been widely used in metric learning (Sohn, 2016; Wu et al., 2018; van den Oord et al., 2018), face recognition (Ranjan et al., 2017; Liu et al., 2017; Wang et al., 2017; 2018; Deng et al., 2019; Zhang et al., 2019), and self-supervised learning (Chen et al., 2020).”
Probabilistic Face Embeddings (2019, methodology): “In particular, we implement the following state-of-the-art loss functions: Softmax+Center Loss [35], Triplet Loss [27], A-Softmax [20] and AM-Softmax [33](9).”