SFace: Sigmoid-Constrained Hypersphere Loss for Robust Face Recognition

Zhong, Yaoyao; Deng, Weihong; Hu, Jiani; Zhao, Dongyue; Li, Xian; Wen, Dongchao

doi:10.1109/tip.2020.3048632

Cited by 93 publications

(43 citation statements)

References 35 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…In this way, stochastic gradient descent optimization can be better controlled. However, related works [57], [43] explore that the term 𝜕𝑑/𝜕𝒙 harbors a hidden modulation function about 𝑑 which would break our intention.…”

Section: Methodsmentioning

confidence: 95%

“…Circle Loss [41] fulfills this purpose with Circle Margin. SFace [57] employs sigmoid functions to mine hard pairs. Meanwhile, SFace also finds that the functions are disturbed in deep back propagation.…”

Section: Related Workmentioning

confidence: 99%

“…Especially, Zhang and Rusinkiewicz [55] have proposed a dynamic modulation for Siamese pairs with CDF, but it is not considerate. Moreover, the deep analysis on back-propagated gradients performed by HyNet [43] and SFace [57] suggests that the hidden modulation deserves more attentions. Additionally, the SFace points out that the strict HEM might be irrational, which indirectly explains the nature of hybrid similarity in HyNet.…”

Section: Related Workmentioning

confidence: 99%

“…While we are formulating a statistic-based dynamic gradient modulation, more details should be explored. Firstly, related works [43], [57] indicate that once a specific metric, e.g., 𝐿 2 distance or inner product is chosen, an implicit modulation harbored in deep back propagation would disorder the elaborate schedule. Secondly, strict HEM should not be encouraged.…”

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

SDGMNet: Statistic-based Dynamic Gradient Modulation for Local Descriptor Learning

Ma¹,

Deng²

2021

Preprint

View full text Add to dashboard Cite

Modifications on triplet loss that rescale the backpropagated gradients of special pairs have made significant progress on local descriptor learning. However, current gradient modulation strategies are mainly static so that they would suffer from changes of training phases or datasets. In this paper, we propose a dynamic gradient modulation, named SDGMNet, to improve triplet loss for local descriptor learning. The core of our method is formulating modulation functions with statistical characteristics which are estimated dynamically. Firstly, we perform deep analysis on back propagation of general tripletbased loss and introduce included angle for distance measure. On this basis, auto-focus modulation is employed to moderate the impact of statistically uncommon individual pairs in stochastic gradient descent optimization; probabilistic margin cuts off the gradients of proportional Siamese pairs that are believed to reach the optimum; power adjustment balances the total weights of negative pairs and positive pairs. Extensive experiments demonstrate that our novel descriptor surpasses previous stateof-the-arts on standard benchmarks including patch verification, matching and retrieval tasks.

show abstract

Section: Methodsmentioning

confidence: 95%

Section: Related Workmentioning

confidence: 99%

Section: Related Workmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

SDGMNet: Statistic-based Dynamic Gradient Modulation for Local Descriptor Learning

Ma¹,

Deng²

2021

Preprint

View full text Add to dashboard Cite

show abstract

“…(4) ResNet100 model trained on MS1MV2 database [10] refined by insightface with [7], (5) ResNet100 model trained on MS1MV2 database [10] with CurricularFace [12], (6) ResNet100 model trained on MS1MV2 database [10] with SFace [19].…”

Section: Baselinementioning

confidence: 99%

MLFW: A Database for Face Recognition on Masked Faces

Wang¹,

Zhong²,

Deng³

2021

Preprint

Self Cite

View full text Add to dashboard Cite

As more and more people begin to wear masks due to current COVID-19 pandemic, existing face recognition systems may encounter severe performance degradation when recognizing masked faces. To figure out the impact of masks on face recognition model, we build a simple but effective tool to generate masked faces from unmasked faces automatically, and construct a new database called Masked LFW (MLFW) based on Cross-Age LFW (CALFW) database. The mask on the masked face generated by our method has good visual consistency with the original face. Moreover, we collect various mask templates, covering most of the common styles appeared in the daily life, to achieve diverse generation effects. Considering realistic scenarios, we design three kinds of combinations of face pairs. The recognition accuracy of SOTA models declines 5%-16% on MLFW database compared with the accuracy on the original images. MLFW database can be viewed and downloaded at http://whdeng.cn/mlfw.

show abstract