CosFace: Large Margin Cosine Loss for Deep Face Recognition

Wang, Hao; Wang, Yitong; Zhou, Zheng; Ji, Xing; Gong, Dihong; Zhou, Jingchao; Li, Zhifeng; Liu, Wei

doi:10.48550/arxiv.1801.09414

Cited by 60 publications

(55 citation statements)

References 21 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The SphereFace [4] is the first method which proposes a multiplicative angular margin penalty to strengthen extra intra-class compactness and inter-class discrepancy simultaneously. Cosface [5] directly adds cosine margin penalty to the target logit and is much easier to train than SphereFace. Arcface [6] proposes an additive angular margin and releases a cleaned version of the MS-celeb-1m which is quite beneficial for face recognition community.…”

Section: Related Workmentioning

confidence: 99%

“…. Specifically, SphereFace [4] fixes m 2 = m 3 = 0, Cosface [5] fixes m 1 = 1, m 2 = 0 and Arcface [6] fixes m 1 = 1, m 3 = 0.…”

Section: Multiface Large Margin Loss Functionsmentioning

confidence: 99%

“…Cosface [5] and Arcface [6] are two margin-based losses that have achieved outstanding performance recently. As shown in Table 1 our methods improve the performance of both Cosface and Arcface with faster convergence speed, which shows the extensive generalization capability.…”

Section: Ablation Study On Lossesmentioning

confidence: 99%

“…Since the traditional softmax Loss lacks an adequate power of discrimination, many novel loss functions are proposed to increase the feature discriminative power in the Euclidean margin, including Center loss [3], Contrastive Loss [1] and Triplet loss [2]. Alternatively, novel angular margin-based losses are proposed, such as Sphereface [4], Cosface [5] and Arcface [6]. These loss functions are more stable and robust due to the intrinsic consistency with softmax and they pay more attention to intra-class compactness while maintaining the property of inter-class separation.…”

Section: Introductionmentioning

confidence: 99%

“…However, most of the existing methods lack careful analysis in the selection of spatial dimension. Typically, the dimension of facial feature is set to one single number by empirical study, e.g., 512 in [3,7,5]. Intuitively, low dimension features lack adequate expressive ability but can be easily optimized, while high dimensional features are hard to optimize but can be more powerful to depict…”

Section: Introductionmentioning

confidence: 99%

See 4 more Smart Citations

MultiFace: A Generic Training Mechanism for Boosting Face Recognition Performance

Xu¹,

Guo²,

Xu³

et al. 2021

Preprint

View full text Add to dashboard Cite

Deep Convolutional Neural Networks (DCNNs) and their variants have been widely used in large scale face recognition(FR) recently. Existing methods have achieved good performance on many FR benchmarks. However, most of them suffer from two major problems. First, these methods converge quite slowly since they optimize the loss functions in a high-dimensional and sparse Gaussian Sphere. Second, the high dimensionality of features, despite the powerful descriptive ability, brings difficulty to the optimization, which may lead to a sub-optimal local optimum. To address these problems, we propose a simple yet efficient training mechanism called MultiFace, where we approximate the original high-dimensional features by the ensemble of low-dimensional features. The proposed mechanism is also generic and can be easily applied to many advanced FR models. Moreover, it brings the benefits of good interpretability to FR models via the clustering effect. In detail, the ensemble of these low-dimensional features can capture complementary yet discriminative information, which can increase the intra-class compactness and inter-class separability. Experimental results show that the proposed mechanism can accelerate 2-3 times with the softmax loss and 1.2-1.5 times with Arcface or Cosface, while achieving state-ofthe-art performances in several benchmark datasets. Especially, the significant improvements on large-scale datasets(e.g., IJB and MageFace) demonstrate the * Equal contributions from both authors. This work was conducted at Tencent.

show abstract

Section: Related Workmentioning

confidence: 99%

“…. Specifically, SphereFace [4] fixes m 2 = m 3 = 0, Cosface [5] fixes m 1 = 1, m 2 = 0 and Arcface [6] fixes m 1 = 1, m 3 = 0.…”

Section: Multiface Large Margin Loss Functionsmentioning

confidence: 99%

Section: Ablation Study On Lossesmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

MultiFace: A Generic Training Mechanism for Boosting Face Recognition Performance

Xu¹,

Guo²,

Xu³

et al. 2021

Preprint

View full text Add to dashboard Cite

show abstract

Improving Generalization via Scalable Neighborhood Component Analysis

Efros

2018

Computer Vision – ECCV 2018

116

128

View full text Add to dashboard Cite

Current major approaches to visual recognition follow an end-to-end formulation that classifies an input image into one of the predetermined set of semantic categories. Parametric softmax classifiers are a common choice for such a closed world with fixed categories, especially when big labeled data is available during training. However, this becomes problematic for open-set scenarios where new categories are encountered with very few examples for learning a generalizable parametric classifier. We adopt a non-parametric approach for visual recognition by optimizing feature embeddings instead of parametric classifiers. We use a deep neural network to learn the visual feature that preserves the neighborhood structure in the semantic space, based on the Neighborhood Component Analysis (NCA) criterion. Limited by its computational bottlenecks, we devise a mechanism to use augmented memory to scale NCA for large datasets and very deep networks. Our experiments deliver not only remarkable performance on ImageNet classification for such a simple nonparametric method, but most importantly a more generalizable feature representation for sub-category discovery and few-shot recognition.

show abstract

Proceedings of the 22nd Engineering Applications of Neural Networks Conference

Symeonidis

Nousi

Tosidis

et al. 2021

Proceedings of the International Neural Networks Society

View full text Add to dashboard Cite

translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. The publisher, the authors and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, expressed or implied, with respect to the material contained herein or for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

show abstract

CosFace: Large Margin Cosine Loss for Deep Face Recognition

Cited by 60 publications

References 21 publications

MultiFace: A Generic Training Mechanism for Boosting Face Recognition Performance

MultiFace: A Generic Training Mechanism for Boosting Face Recognition Performance

Improving Generalization via Scalable Neighborhood Component Analysis

Proceedings of the 22nd Engineering Applications of Neural Networks Conference

Contact Info

Product

Resources

About