Proxy Attention

• Let $I_s\in {\mathbb{R}}^{W\times H\times C}$ be a random source image
• applied saliency map $I_{vs}$
  • $I_{vs}=f(I_s)\in {\mathbb{R}}^{W\times H}$
  • Augmented image $I_a$
  • $\odot$ elementwise multi
  • Mask $M\in \{0,1{\}}^{W\times H}$