Algorithm 2. Training procedure of ASNet

Input: Original input frame clips $x^g$

Output: ${\theta }_{cnn}$ and ${\theta }_{spa}$

1: Initialize ${\theta }_{cnn}$ and ${\theta }_{spa}$

2: while $k\le K$ do

3:   Get $x^c$ through conventional cropping on $x^g$

4:   Get $x^s$ through SPA on $x^g$

5:   Take $x^c$ and $x^s$ as inputs; Fix ${\theta }_{spa}$; Train ${\theta }_{cnn}$

6:   Fix ${\theta }_{cnn}$; Train ${\theta }_{spa}$ through Algorithm 1

7: end while