Convolutional Block Attention Module (CBAM)

https://www.youtube.com/watch?v=lpFufRf9jFI

• the main idea is this:
  • if you give the model the ability to learn that specific areas are more important than others, then hopefully we’ll get a better result
• the authors realized that applying a channel attention module before spatial attention modules works the best
  • no intuition, just emperical trials
• How the channel attention module works:
  • 
  • 
    • first do avg pool and max pool to get representations of layers
      • this is across the spatial dimension (so for channel Red, what is the average value)
  • 
    • now we pass it into a shared MLP that outputs two vectors, one for avg pool and the other for maxpool
      • this shared mlp is exactly like what we had in squeeze and excite
        • the image is prob misleading since it hsould be two different MLPs (since in the math formula, they are separate)
    • but ultimately, we need a single value that determines which channel is more important than others, so we sum it all and use a sigmoid
    • 
  • this feels like Squeeze-and-Excitation layer, but there’s an added maxpool (so we get more information about individual channels before the shared MLP)
• How the spatial attention module works:
  • 
  • first apply avg pool and max pool
    • but unlike before, we’re doing these avg pool per pixel (across channels)
      • so for each pixel vector, we AVGPool all of the channels for that individual pixel
  • 
    • now we concat and put it through a conv layer to learn which region is more important
    • they used a 7x7 kernel cause it performed the best
  • finally, just put it through a sigmoid
    • 
• We integrate these layers just like we integrate Squeeze-and-Excitation layer
  •