Skip to main content

Table 4 Verification accuracy (%) of different state-of-the-art methods on RFW. BiometricNet+ outperforms the state of the art on all the considered datasets. The numbers in brackets for the BiometricNet+ entries represent the verification accuracy without the four flips at inference time

From: Cancelable templates for secure face verification based on deep learning and random projections

Method

Caucasian

Indian

Asian

African

CenterLoss [72]

87.18

81.92

79.32

78.00

SphereFace [46]

90.80

87.02

82.95

82.28

VGGFace2 [78]

89.90

86.13

84.93

83.38

ArcFace [48]

97.37

95.68

94.55

93.87

CosFace [47]

96.63

94.68

93.50

92.17

IMAN-A [79]

-

94.15

91.15

91.42

RL-RBN [80]

97.08

95.63

95.57

94.87

BiometricNet+ NOCB

99.33

98.75

98.33

98.12

(99.12)

(98.41)

(98.13)

(97.98)