Table 1.

The details of the ESA-rPPGNet network structure. Input is the input dimension of each layer $H\times W\times T\times C$, Operator is the convolution operation in each layer and the corresponding convolution kernel size, and Exp size is the channel of DWC. Opt size is the final output channel number of the Block, Stride is the step size of the DWC, SA indicates whether to use the 3D-SA module, where 1 indicates yes, 0 indicates no, and NL indicates the nonlinear layer type: 0 means ReLU6, 1 means $h_{-}swish$.

Input	Operator	Exp Size	Opt Size	Stride	SA	NL
${128}^2\times 128\times 3$	Conv3d, 3 × 3 × 3	-	16	1 × 2 × 2	-	-
${64}^2\times 128\times 16$	EBlock, 3 × 3 × 3	16	16	1 × 2 × 2	1	0
${32}^2\times 128\times 16$	EBlock, 3 × 3 × 3	72	24	1 × 2 × 2	0	0
${16}^2\times 128\times 24$	EBlock, 3 × 3 × 3	88	24	1 × 1 × 1	0	0
${16}^2\times 128\times 24$	EBlock, 5 × 5 × 5	96	40	2 × 2 × 2	1	1
$8^2\times 64\times 40$	EBlock, 5 × 5 × 5	240	40	1 × 1 × 1	1	1
$8^2\times 64\times 40$	EBlock, 5 × 5 × 5	240	40	1 × 1 × 1	1	1
$8^2\times 64\times 40$	EBlock, 5 × 5 × 5	120	48	1 × 1 × 1	1	1
$8^2\times 64\times 48$	EBlock, 5 × 5 × 5	144	48	1 × 1 × 1	1	1
$8^2\times 64\times 48$	EBlock, 5 × 5 × 5	288	96	2 × 2 × 2	1	1
$4^2\times 32\times 96$	EBlock, 5 × 5 × 5	576	96	1 × 1 × 1	1	1
$4^2\times 32\times 96$	EBlock, 5 × 5 × 5	576	96	1 × 1 × 1	1	1
$4^2\times 32\times 96$	Conv3d, 1 × 1 × 1	-	576	1 × 1 × 1	-	-
$4^2\times 32\times 576$	DCBlock, 4 × 1 × 1	-	288	2 × 1 × 1	-	-
$4^2\times 64\times 288$	DCBlock, 4 × 1 × 1	-	144	2 × 1 × 1	-	-
$4^2\times 128\times 144$	ConvGRU	-	64	-	-	-
$4^2\times 128\times 64$	GAP	-	64	-	-	-
$1^2\times 128\times 64$	Conv3d, 1 × 1 × 1	-	1	1 × 1 × 1	-	-