Table 2.

Accuracy rates for gender prediction on UTKFace [83], for fully visible faces (denoted by ), lower-half-visible faces (denoted by ) and upper-half-visible faces (denoted by )

Method	Train faces	Test faces	Accuracy (%)
ResNet-50+PyNADA [19]			90.80
Teacher VGG-f			92.78
Teacher VGG-face			92.20
Teacher ResNet-50			90.88
Teacher VGG-f			78.13
Teacher VGG-face			73.05
Teacher ResNet-50			72.69
Teacher VGG-f			85.69
Teacher VGG-face			88.18
Teacher ResNet-50			83.20
VGG-f			88.70
VGG-face			90.62
ResNet-50			86.47
VGG-f			88.92
VGG-face			88.26
ResNet-50			88.75
VGG-f (standard T-S)	+		89.13
VGG-face (standard T-S)	+		88.45
ResNet-50 (standard T-S)	+		$89.{45}^{†}$
VGG-f (triplet loss T-S)	+		$89.{55}^{†}$
VGG-face (triplet loss T-S)	+		88.31
ResNet-50 (triplet loss T-S)	+		89.19
VGG-f (triplet loss + standard T-S)	+		$89.{82}^{†}$
VGG-face (triplet loss + standard T-S)	+		$90.{35}^{†}$
ResNet-50 (triplet loss + standard T-S)	+		$89.{63}^{†}$

A state-of-the-art model [19] is included as reference. The results of distilled models that are significantly better than the student trained on upper-half-visible faces, according to a paired McNemar’s test [9] at a significance level of 0.05, are marked with $†$