gather-excite

**Gather-Excite (GE)** is a **spatial attention mechanism that gathers local spatial context and then excites (modulates) feature responses** — extending the squeeze-and-excitation concept from channel attention to spatial attention by gathering spatial neighborhoods. **How Does Gather-Excite Work?** - **Gather**: Aggregate spatial context at multiple scales using depth-wise convolutions or average pooling at different resolutions. - **Excite**: Use the gathered context to produce spatial attention weights. - **Modulate**: Multiply feature maps by the spatial attention weights. - **Variants**: GE-θ (parameterized gather), GE-θ+ (with skip), GE-θ- (lightweight). - **Paper**: Hu et al. (2018). **Why It Matters** - **Spatial SE**: Extends the highly successful SE concept to the spatial dimension. - **Multi-Scale**: The gathering operation captures context at multiple spatial scales. - **Complementary**: Can be combined with channel attention (SE) for full channel+spatial attention. **Gather-Excite** is **spatial context for feature modulation** — gathering neighborhood information to tell each location how important it is.