cutmix

CutMix is a regional augmentation that cuts a rectangular patch out of one image and pastes it onto another, then sets the label to the area-weighted mixture of the two classes. If the pasted patch covers 30 percent of the canvas, the target becomes 0.7 of the background image's class and 0.3 of the patch's class. It can be seen as a fusion of two earlier ideas, Cutout (deleting a region) and Mixup (interpolating labels), but instead of zeroing the region or blending pixels translucently, CutMix fills the hole with real content from another image.

The key advantage over Mixup is that every pixel comes from a real image, so there are no ghostly semi-transparent blends; the cut-and-paste forces the network to recognise an object from partial views and to attend to multiple regions, which improves localisation as well as classification. The area-proportional label is what keeps the supervision honest: the model is rewarded for distributing its confidence in proportion to how much of each object is actually visible.

CutMix consistently improves ImageNet accuracy and robustness and is a staple of modern training recipes, frequently applied stochastically alongside Mixup so that each batch randomly receives one augmentation or the other. Its hyperparameters are the Beta distribution controlling patch size and the probability of applying it per batch.