Table 3 Verification accuracy (%) of different methods on LFW, YTF, CALFW, CPLFW, CFP, and AgeDB. BiometricNet+ outperforms the state of the art on all the considered datasets. The # Images column indicates the dimension of the training set for each approach. The numbers in brackets for the BiometricNet+ entries represent the verification accuracy obtained without the four flips at inference time

DeepID [10]

0.2M

99.47

93.20

-

-

-

-

SphereFace [46]

0.5M

99.42

95.0

90.30

81.40

94.38

91.70

SphereFace+ [71]

0.5M

99.47

-

-

-

-

-

CenterLoss [72]

0.7M

99.28

94.9

85.48

77.48

-

-

Baidu [73]

1.3M

99.13

-

-

-

-

-

UniformFace [74]

2.21M

99.80

97.70

-

-

-

-

VGGFace [11]

2.6M

98.95

97.30

90.57

84.00

-

-

MarginalLoss [75]

3.8M

99.48

95.98

-

-

-

-

DeepFace [76]

4.4M

97.35

91.4

-

-

-

-

RangeLoss [77]

5M

99.52

93.7

-

-

-

-

CosFace [47]

5M

99.73

97.6

-

-

95.44

-

ArcFace [48]

5.8M

99.82

98.02

95.45

92.08

98.37

95.15

FaceNet [14]

200M

99.63

95.10

-

-

-

89.98

BiometricNet [16]

3.8M

99.80

98.06

97.07

95.60

99.35

96.12

BiometricNet+ r=64

4.4M

99.28

97.08

96.17

94.18

97.73

94.45

(99.11)

(96.98)

(96.03)

(94.02)

(97.59)

(94.21)

BiometricNet+ r=128

4.4M

99.62

97.78

96.67

95.33

99.13

95.57

(99.32)

(97.39)

(96.29)

(95.11)

(99.01)

(95.27)

BiometricNet+ r=256

4.4M

99.72

97.85

96.70

95.67

99.18

95.82

(99.41)

(97.56)

(96.41)

(95.27)

(99.04)

(95.43)

BiometricNet+ NOCB

4.4M

99.82

98.16

97.10

95.85

99.43

96.24

(99.75)

(98.01)

(96.88)

(95.30)

(99.37)

(96.17)