--- title: "CBAM: Convolutional Block Attention Module" --- # CBAM: Convolutional Block Attention Module(ECCV2018) 這篇主要結合 Spatial & Channel Domain 的 Attention ### CBAM Block ![](https://i.imgur.com/dPEHOFs.png) ### **Channel attention** ![](https://i.imgur.com/ESO2Pis.png) ![](https://i.imgur.com/vOmfa4X.png) 跟SE-Net一樣用兩個FC後cancat在一起做sigmoid取得scale $r$一樣是16 #### 針對 Channel 的實驗 實驗證明Avg $+$ Max會比SE提出的單項pooling好 ![](https://i.imgur.com/tuYm5jC.png) ### **Spatial attention** ![](https://i.imgur.com/QjWScdo.png) 將$F'$的channel(只有只對axis=[3])進行Maxpool跟Avgpool成兩個二維的Map再concat在一起做conv(7 $*$ 7) 最後sigmoid成第二個scale #### 針對 Spatial 的實驗 ![](https://i.imgur.com/XDNgJmI.png) ### Attention Module 的順序影響 ![](https://i.imgur.com/crKlWxR.png) ### Classification results on ImageNet-1K ![](https://i.imgur.com/TkNo9FX.png) ### Grad-CAM visualization ![](https://i.imgur.com/ecGLzVW.jpg) ![](https://i.imgur.com/Bp0xor1.jpg)