Skip to main content

Table 4 Important milestones in face recognition with corresponding LFW verification accuracies

From: Deep-learned faces: a survey

Year

Publication

DCNN architecture

Loss function

Train data

LFW (%)

2014

DeepFace [27]

9 layer deep CNN

Softmax

Private dataset

97.35

2014

DeepId [54]

9 layer deep CNN

Softmax

CelebFaces+

97.45

2014

DeepID2 [55]

9 layer deep CNN

Softmax, contrastive

CelebFaces [106]

99.15

2014

DeepID2+ [56]

9 layer deep CNN

Softmax, contrastive

CelebFaces+,

99.47

    

WDRef [129]

99.53

2015

VGGFace [37]

VGGNet [35]

Softmax, triplet

VGGFace [37]

98.95

2015

DeepID3 [28]

GoogleNet [36],

Softmax, contrastive

CelebFaces+,

99.53

  

VGGNet [35]

 

WDRef [129]

99.53

2015

FaceNet [29]

GoogleNet [36],

Triplet

Private dataset [29]

99.63

  

Zeiler_Fergus [125]

   

2015

Baidu [57]

DCNN with 9 convolutions

Triplet

Private dataset

99.77

2018

SphereFace [60]

ResNet-64 [70]

SphereFace [60]

CASIA-WebFace [88]

99.42

2018

CosFace [61]

ResNet-64 [70]

CosFace [61]

CASIA-WebFace [88]

99.33

    

Private dataset

99.73

2019

ArcFace [38]

ResNet-100 [70]

ArcFace [38]

ms1m

99.83

Open Source Implementations

-

DLIB library

ResNet-34 [70]

Triplet

VGGFace [37],

99.38

    

FaceScrub [108]

 

2016

OpenFace [65]

GoogleNet [36]

Triplet

CASIA-WebFace [88],

92.92

    

FaceScrub [108]

 

-

FaceNet_Re [66]

Inception-ResNet-v1 [71]

Softmax

VGGFace2 [78]

99.65

  

Inception-ResNet-v1 [71]

Softmax

CASIA-WebFace [88]

99.05

  

Inception-ResNet-v1 [71]

Centre [79]

VGGFace2 [78]

99.2