CutMix is a data augmentation technique that cuts a rectangular region from one image and pastes it onto another — mixing the labels proportionally to the area of the cut region, combining the benefits of Cutout (occlusion robustness) and Mixup (label smoothing).

How Does CutMix Work?

• Sample $lambda$: $lambda sim ext{Beta}(alpha, alpha)$.
• Cut Region: Random box with area ratio $1 - lambda$ of the total image.
• Paste: Replace the cut region in image $A$ with the corresponding region from image $B$.
• Labels: $ ilde{y} = lambda y_A + (1-lambda) y_B$ (proportional to visible area).
• Paper: Yun et al. (2019).

Why It Matters

• Best of Both: Unlike Mixup (blurry blends) or Cutout (wasted pixels), CutMix uses all pixel information.
• Localization: Forces the model to learn from local regions, improving weakly-supervised localization.
• SOTA: Widely adopted in modern ImageNet training recipes alongside Mixup and RandAugment.

CutMix is a surgical transplant between images — cutting and pasting regions to create informative training samples that use every pixel.