snapmix

**SnapMix** is a **semantically proportional data augmentation technique for fine-grained recognition** — using Class Activation Maps (CAM) to determine the semantic importance of each pixel, then mixing images with labels weighted by the semantic content, not just area. **How Does SnapMix Work?** - **CAM**: Compute Class Activation Maps for both images using the current model. - **Cut Region**: Select a region from image $B$ (random or guided by CAM). - **Paste**: Replace the corresponding region in image $A$. - **Semantic Labels**: Weight labels by the CAM activation in the visible regions, not by pixel area. - **Paper**: Huang et al. (2021). **Why It Matters** - **Fine-Grained**: Designed for tasks where small, discriminative parts define the class (bird species, car models). - **Semantic Labeling**: CAM-based label mixing is more accurate than area-based (a 50% area overlap may contain 90% of the discriminative features). - **Better Than CutMix**: Significant improvements on fine-grained benchmarks (CUB-200, Stanford Cars). **SnapMix** is **semantically-aware CutMix** — using attention maps to assign honest labels that reflect the true informative content of mixed images.