Figure 4:

The proposed attention-driven fusion block in the i^th layer. (a) Overview of the fusion block, which contains four attention mechanisms: hard attention, illustrated in (b) and the output features are denoted as $f_i^{u_1}$; soft attention, illustrated in (c) and the output features are denoted as $f_i^{u_2}$; and two mutual attentions, illustrated in (d) and the output features are denoted as $f_i^{u_3}$ and $f_i^{u_4}$.