Table 3.

Network structure of the facial landmark feature module.

	Unit	Layer	Filter/Stride	Output Size
Input	0			112 × 112 × 1
Facial Landmark Feature Network	1	Conv-BN-ReLU Conv-BN-ReLU Conv-BN-ReLU	7 × 7, 64/2 7 × 7, 128/2 7 × 7, 256/2	53 × 53 × 64 24 × 24 × 12 89 × 9 × 256
Output	2	GlobalAvgPool		256

BN: batch normalization. The stride is 2, but the feature map size is reduced by more than 0.5-times because padding is not performed during convolution.