CBAM Paper Walkthrough: The Double

The Convolutional Block Attention Module (CBAM) enhances feature quality by integrating two attention mechanisms: Channel Attention Module (CAM) and Spatial Attention Module (SAM). CAM operates by applying global max-pooling and average-pooling across the spatial dimensions, followed by a shared Multi-Layer Perceptron (MLP) to learn channel importance, which is then modulated by a sigmoid function. SAM differentiates itself by pooling across the channel dimension for each spatial location using max and average pooling, followed by a $7 \times 7$ convolution layer. The module is designed to be modular, maintaining the input tensor's shape, allowing it to be integrated into various backbone models like ResNeXt without requiring shape alteration. Experimental results indicated that CBAM generally outperformed baseline models, including SENet, and demonstrated parameter and FLOP reduction compared to deeper models, suggesting suitability for low-power devices. Qualitative evaluation using Grad-CAM showed that CBAM refines focus, leading to more accurate object localization by emphasizing relevant spatial regions over general feature extraction.

The evolution from channel attention (SENet) to the dual attention approach of CBAM, incorporating both spatial and channel weighting, suggests a necessary shift in how attention is modeled in vision tasks. The decision to use global max-pooling and average-pooling separately in CAM, and then transition to channel pooling in SAM with a subsequent convolution layer, reflects an attempt to capture complementary forms of relational information: macro-level channel importance versus fine-grained spatial context. The empirical evidence demonstrating that the combined attention mechanism results in superior performance and efficiency implies that simply assigning weights based on one dimension is insufficient; true feature refinement requires holistic spatial-channel correlation. The pattern where sequential application optimizes performance suggests a dependency in the hierarchical nature of visual information processing, where understanding channel relevance precedes precise spatial localization. The reduction in parameters relative to deeper models highlights a tension: achieving state-of-the-art accuracy while adhering to constraints for deployment on low-power systems, which forces a re-evaluation of what constitutes an "optimal" feature representation versus computational feasibility. What is the cost of this efficiency gain when moving from learned global features to explicit spatial weighting?