Course contentsShow
← Machine Learning and Deep Learning
Lesson 770 of 3,53818. Regularization Techniques for Deep NetworksFree lesson

CutMix: Regional Mixing Strategy

Combining image patches from different examples while adjusting labels proportionally to regions.

CutMix: Regional Mixing Strategy

You'll learn how CutMix creates hybrid training examples by cutting and pasting image regions between samples while adjusting labels proportionally.

What Is CutMix?

CutMix is a data augmentation technique that combines regional patches from two different training images. Unlike Mixup (which blends entire images together), CutMix cuts out a rectangular region from one image and pastes it onto another. The training label becomes a weighted mix of both original labels, proportional to the pixel area each image contributes.

The Core Idea

Think of CutMix like creating a photo collage: you take a chunk from photo A, paste it over photo B, and then tell the model "this hybrid is partially A and partially B" based on how much space each occupies.

The process:

  1. Select two training images with labels
  2. Randomly generate a bounding box (position and size)
  3. Cut the region from image A and paste it onto image B
  4. Adjust the label: if the patch covers 30% of the area, the new label is 0.3 × label_A + 0.7 × label_B

Why CutMix Works

CutMix forces the network to recognize objects from partial views and learn from local features rather than relying on the full context. It also prevents the network from focusing only on the most discriminative region, improving robustness. The sharp boundaries (unlike Mixup's blending) preserve spatial information better for convolutional networks.

Unlike Cutout (which replaces regions with zeros), CutMix fills the space with meaningful content from another example, making more efficient use of every pixel.

Key Takeaway: CutMix creates hybrid training examples by swapping rectangular regions between images and mixing labels proportionally to patch size, helping networks learn robust local features while preventing overfitting.