Coordinate Attention is a lightweight attention mechanism that encodes channel relationships and long-range spatial dependencies — by decomposing global pooling into two 1D operations (horizontal and vertical), preserving positional information that SE-Net's global average pooling discards.

How Does Coordinate Attention Work?

• Horizontal Pool: Average pool along the width dimension -> $z_h(h) in mathbb{R}^{C imes H imes 1}$.
• Vertical Pool: Average pool along the height dimension -> $z_w(w) in mathbb{R}^{C imes 1 imes W}$.
• Transform: Concatenate, pass through shared 1×1 conv + BN + activation, then split.
• Attention: Sigmoid-activated 1×1 conv produces 2D spatial-aware channel attention maps.
• Paper: Hou et al. (2021).

Why It Matters

• Position-Aware: Unlike SE (global avg pool -> loses position), Coordinate Attention preserves spatial structure.
• Lightweight: Minimal additional parameters and FLOPs.
• Object Detection: Particularly effective for dense prediction tasks where spatial position matters.

Coordinate Attention is SE with spatial awareness — encoding directional position information into channel attention for better localization.