Selective Kernel Networks

URL

https://arxiv.org/pdf/1903.06586.pdf

TL;DR

• SKNet 给 N 个不同感受野分支的 feature 通道赋予权重，结合了 Attention to channel 和 select kernel

SKNet网络结构

数学表达

$X\in\mathbb R^{H'\times W' \times C'} \overset{\tilde F}{\longrightarrow} \tilde U \in \mathbb R^{H\times W\times C}$

$X\in\mathbb R^{H'\times W' \times C'} \overset{\hat F}{\longrightarrow} \hat U \in \mathbb R^{H\times W\times C}$

$U=\tilde U + \hat U$

$s_c = F_{gp}(U_c) = \frac{1}{H\times W}\sum_{i=1}^H\sum_{j=1}^W U_c(i, j)$

$z = F_{fc}(s) = \delta(\beta (Ws)),\ \ \ \ W\in\mathbb R^{d\times C},\ \ \ \ d = max(\frac{C}{r}, L)$

$a_c = \frac{e^{A_cz}}{e^{A_cz} + e^{B_cz}},\ \ b_c = \frac{e^{B_cz}}{e^{A_cz} + e^{B_cz}},\ \ \ \ A_c,B_c\in\mathbb R^{1\times d}$

$V_c = a_c . \tilde U_c + b_c . \hat U_c,\ \ \ \ V_c\in\mathbb R^{H\times W}$

SKNet实验结果

• ImageNet

• other

Thoughts

• SENet 与 SKNet 属于 Attention to channel，ULSAM 属于 Attention to HW，两个合起来是否可以替代 Non-local——在 THW上的 Attention