A Real-Time and Long-Term Face Tracking Method Using Convolutional Neural Network and Optical Flow in IoT-Based Multimedia Communication Systems

Table 1

Architectures of cascade convolutional networks. “” stands for kernel size, “” means stride, and “” is padding number.

Stage 1
Input	Color-scale image
Convolution	Outputs: 10, : , : 1, : 1
MaxPooling	: , : 2
Convolution	Outputs: 16, : , : 1, : 1
Convolution	Outputs: 32, : , : 1, : 1
Convolution	Outputs: 6, : , : 1, : 1
Stage 2
Input	Color-scale image
Convolution	Outputs: 28, : , : 1, : 1
MaxPooling	: , : 2
Convolution	Outputs: 48, : , : 1, : 1
MaxPooling	: , : 2
Convolution	Outputs: 64, : , : 1, : 1
Dense	Outputs: 128
Dense	Outputs: 6
Stage 3
Input	Color-scale image
Convolution	Outputs: 32, : , : 1, : 1
MaxPooling	: , : 2
Convolution	Outputs: 64, : , : 1, : 1
MaxPooling	: , : 2
Convolution	Outputs: 128, : , : 1, : 1
Dense	Outputs: 256
Dense	Outputs: 6