CutMix
Data augmentation technique that cuts out a rectangular region from one image and replaces it with a region from another image.
CutMix cuts a patch from one image and replaces it with a patch from another – labels are adjusted proportionally. Forces more robust feature usage than Mixup.
Explanation
Labels are adjusted proportionally to the area. Forces the model to use all image regions for classification, not just a dominant area.
Marketing Relevance
CutMix is standard in modern computer vision training and often better than Cutout or Mixup alone.
Common Pitfalls
Patch size must be well chosen. With small objects, important information can be lost.
Origin & History
Introduced in 2019 by Yun et al. (KAIST). Combines ideas from Cutout (2017) and Mixup (2017) and achieved state-of-the-art on ImageNet and CIFAR.
Comparisons & Differences
CutMix vs. Mixup
Mixup blends globally; CutMix replaces locally. CutMix preserves local pixel statistics and trains more robust local features.
CutMix vs. Cutout
Cutout masks a region with zeros (information is lost); CutMix replaces it with useful information from another image.