CutMix

• @yunCutMixRegularizationStrategy2019
• images are augmented by sampling patch coordinates, x, y, h, w from a uniform distribution
• selected patch is replaced at the
• corresponding location with a patch from the other randomly picked image from the current mini-batch during training.
• M is the image mask, xa and xb are images, λ is the proportion of label, and ya and yb are the labels of images.

x_{new}= M.x_{a}+(1-M).x_{b}

- $$ y_{new}= \lambda.y_{a}+ (1-\lambda).y_{b}