CutMix: Regularization Strategy to Train Strong Classifiers With Localizable Features

This paper proposes CutMix, a data augmentation strategy for training strong image classifiers with localizable features. Prior regional dropout methods improve generalization and object localization by making the network attend to less discriminative object parts, but they remove informative pixels by overlaying patches of black pixels or random noise, which causes information loss and makes training inefficient. CutMix instead cuts patches and pastes them among training images, and crucially mixes the ground-truth labels in proportion to the area of the patches, so that no pixels are wasted while the regularization effect of regional dropout is retained.

CutMix consistently outperforms state-of-the-art augmentation strategies on CIFAR and ImageNet classification as well as on the ImageNet weakly-supervised localization task. Beyond classification, a CutMix-trained ImageNet classifier used as a pretrained model gives consistent gains on Pascal object detection and MS-COCO image captioning benchmarks. The authors further show that CutMix improves model robustness against input corruptions and enhances out-of-distribution detection, making it broadly useful across tasks.