SnapMix

• @huangSnapMixSemanticallyProportional2021
• augments training images by ex- tracting and merging random image regions of dif- ferent sizes, where the region size is drawn through the beta distribution for both the images.
• Gen- erated image label is assigned based on semantic composition from normalized (sum to one) CAMs.
• However, the summation of label coefficients can exceed beyond one depending on the semantic composition of the output image.