Regional dropout strategies have been proposed to enhance the performance of convolutional neural network classifiers. They have proved to be effective for guiding the model to attend on less discriminative parts of objects, thereby letting the network generalize better and have better object localization capabilities. However, the deleted regions are usually zeroed-out or filled with random noise, greatly reducing the proportion of informative pixels on training images. We propose CutMix, where patches are cut and pasted among training images where the ground truth labels are also mixed proportionally to the area of the patches. CutMix makes efficient use of training pixels and retains the regularization effect of regional dropout, achieving state-of-the-art results on CIFAR and ImageNet classification, weakly-supervised localization, and transfer learning tasks.

Deep convolutional neural networks (CNNs) have shown promising performances on various computer vision problems such as image classification, object detection, semantic segmentation, and video analysis. To prevent overfitting, "random feature removal regularizations" have been proposed, including dropout and regional dropout methods like Cutout. While effective for regularization, deleted regions are usually zeroed-out or filled with random noise, greatly reducing the proportion of informative pixels on training images. This means the model is trained with "no uninformative pixel" being addressed — wasting valuable training signal.

CutMix addresses this limitation by replacing removed regions with patches from other training images, ensuring "no uninformative pixel during training" while maintaining the regularization benefits of regional dropout. The ground truth labels are mixed proportionally to the area of the combined patches. This simple yet effective strategy simultaneously improves classification accuracy, localization ability, model robustness, and uncertainty estimation, demonstrating that efficient use of every training pixel matters.

Regional Dropout: Methods removing random regions in images have been proposed to enhance generalization. Cutout randomly masks out square regions, while Random Erasing fills them with random values. CutMix differs fundamentally by "filling with patches from another training image" rather than zeroing or noise-filling. Mixup: CutMix shares similarity with Mixup in that both combine two samples with linearly interpolated labels. However, Mixup blends entire images, producing locally ambiguous and unnatural samples. "CutMix overcomes the problem by replacing the image region with a patch from another training image," producing more locally natural results. CutMix is also "complementary to the above methods because it operates on the data level, without changing internal representations or architecture."

The CutMix combining operation generates new training samples as: x_tilde = M * x_A + (1 - M) * x_B and y_tilde = lambda * y_A + (1 - lambda) * y_B, where M is a binary mask indicating the cut region, and lambda is sampled from Beta(alpha, alpha). The rectangular mask dimensions follow: r_w = W * sqrt(1 - lambda) and r_h = H * sqrt(1 - lambda), making the cropped area ratio r_w * r_h / (W * H) = 1 - lambda. This ensures that the label mixing ratio exactly corresponds to the pixel area ratio. Implementation is straightforward: "CutMix is simple and incurs a negligible computational overhead" compared to existing data augmentation techniques.

The authors verify that CutMix effectively forces the model to learn object recognition from partial views. Class Activation Mapping (CAM) visualizations reveal that "CutMix can take advantage of the mixed region on image, but Cutout cannot" — because CutMix provides meaningful content in the replaced region, the model learns to recognize objects from their visible parts while also extracting information from the pasted region. The method achieves three desirable properties simultaneously: "usage of full image region," "regional dropout" effect, and "mixed image & label" correspondence.

ImageNet Classification: CutMix achieves 21.40% top-1 error, representing a +2.28% improvement over the ResNet-50 baseline, outperforming all considered augmentation strategies. Notably, "CutMix improves the performance by +2.28% while increased depth (ResNet-50 to ResNet-152) boosts +1.99%," demonstrating that augmentation benefits can exceed architectural improvements. On CIFAR-100 with PyramidNet-200, CutMix achieves 14.47% top-1 error, +1.98% higher than the baseline.

Weakly Supervised Object Localization: CutMix outperforms Mixup on localization accuracies by +5.51% and +1.41% on CUB200-2011 and ImageNet, respectively. CutMix also achieves comparable localization accuracies to dedicated state-of-the-art WSOL methods. Robustness and Uncertainty: "CutMix significantly improves the robustness to adversarial attacks" compared to other augmentation methods, and "significantly alleviates the over-confidence of the model" on out-of-distribution detection. Transfer Learning: CutMix outperforms Mixup and Cutout on Pascal VOC detection and image captioning benchmarks.

CutMix provides a simple and effective training strategy for training CNNs with strong classification and localization ability. By replacing removed regions with informative patches from other images and mixing labels proportionally, the method achieves "no extra cost" while delivering consistent improvements across classification, localization, transfer learning, robustness, and uncertainty quantification tasks. The approach is complementary to architectural advances and other training strategies, making it a practical and broadly applicable tool for improving deep learning models.