This article reviews and implements the CBAM (Convolutional Block Attention Module) paper by Woo et al. [1]. Released in 2018, CBAM enhances CNN feature quality through a dual attention mechanism, predating the Vision Transformer (ViT) introduced in 2020. It remains relevant due to its lightweight design, making it suitable for low-power deployment.

CBAM improves upon SENet (Squeeze-and-Excitation Network) by adding spatial attention. While SENet applies attention only across channels, CBAM introduces two sub-modules: the Channel Attention Module (CAM) and the Spatial Attention Module (SAM). CAM uses both global average-pooling and global max-pooling, followed by a shared MLP with a reduction ratio, to generate channel weights. SAM applies max and average pooling across channels, then a convolution layer to produce a spatial attention map.

The refined output retains the input tensor dimensions (C×H×W), allowing seamless integration into any backbone architecture. The implementation is demonstrated using PyTorch, providing a practical guide for applying this double-attention mechanism to computer vision tasks.