Table 3.

The specific architecture of the PMB-SCN based on SlowFast.

Stage	Output Sizes	Slow Pathway	Fast Pathway
raw clip	64 × 24 × 224	-
data layer	Slow: 8 × 224 × 224	stride 8 × 1 × 1	stride 2 × 1 × 1
	Fast: 32 × 224 × 224
conv1	Slow: 8 × 112 × 112	1 × 7 × 7, 64	5 × 7 × 7, 8
	Fast: 32 × 112 × 112	stride 1 × 2 × 2	stride 1 × 2 × 2
pool1	Slow: 8 × 56 × 56	1 × 3 × 3, max
	Fast: 32 × 56 ×56	stride, 1 × 2 × 2
res2	Slow: 8 × 56 × 56	$\left[\begin{array}{c}1\times 3\times 3, 128\\ 1\times 3\times 3, 128\end{array}\right]\times 3$	$\left[\begin{array}{c}3\times 3\times 3, 16\\ 1\times 3\times 3, 16\end{array}\right]\times 3$
	Fast: 32 × 56 × 56
res3	Slow: 8 × 28 × 28	$\left[\begin{array}{c}1\times 3\times 3, 256\\ 1\times 3\times 3, 256\end{array}\right]\times 4$	$\left[\begin{array}{c}3\times 3\times 3, 32\\ 1\times 3\times 3, 32\end{array}\right]\times 4$
	Fast: 32 × 28 × 28
res4	Slow: 8 × 14 × 14	$\left[\begin{array}{c}3\times 3\times 3, 512\\ 1\times 3\times 3, 512\end{array}\right]\times 6$	$\left[\begin{array}{c}3\times 3\times 3, 64\\ 1\times 3\times 3, 64\end{array}\right]\times 6$
	Fast: 32 × 14 × 14
res5	Slow: 8 × 7 × 7	$\left[\begin{array}{c}3\times 3\times 3, 1024\\ 1\times 3\times 3, 1024\end{array}\right]\times 3$	$\left[\begin{array}{c}3\times 3\times 3, 128\\ 1\times 3\times 3, 128\end{array}\right]\times 3$
	Fast: 32 × 7 × 7
	1 × 1 × 1	Spatiotemporal pooling, concatenate fc layer with softmax