Cut and Mix

• @yunCutMixRegularizationStrategy2019
• instead of deleting a patch, the patch is replaced with some other image region
• y this approach, an image shares multiple class labels, whereas the major class label belongs to the original class label
• Hence, the model learns to differentiate between two classes within a single image.
• CutMix can be defined by the following operations $\stackrel{\sim }{x}=M\odot x_A+(1-M)\odot x_B$
• $\stackrel{\sim }{y}=\lambda y_A+(1-\lambda )y_B$
• where $x$ is an RGB image, $y$ is the respective label, $M$ is a binary mask of the patch of the image that will be dropped and $\odot$ represents element wise multiplication. The new training sample $\stackrel{\sim }{x},\stackrel{\sim }{y}$ is created by combining two other training samples $x_A,y_A$ and $x_B,y_B$. To control the combination ratio $\lambda$, a sample from the $\beta (1,1)$ distribution is chosen. This combination is quite similar to Mixup.