CBAM: Convolutional Block Attention Module

May 2020

tl;dr: Improvement over SENet.

Overall impression

Channel attention module is very much like SENet but more concise. Spatial attention module concatenates mean pooling and max pooling across channels and blends them together.

Each attention is then used sequentially with each feature map.

The Spatial attention module is modified in Yolov4 to a point wise operation.

Key ideas

Technical details