CutMix is a novel data augmentation technique designed to enhance the performance of convolutional neural networks (CNNs) by improving both classification and localization capabilities. Unlike traditional regional dropout methods that remove informative pixels, CutMix replaces these regions with patches from other training images, maintaining the regularization effect while ensuring efficient use of training pixels. This approach allows the model to focus on less discriminative parts of objects, improving generalization and localization accuracy. The key contributions of CutMix include: 1. **Efficient Training**: By replacing removed regions with patches, CutMix avoids information loss and maintains training efficiency. 2. **Superior Performance**: CutMix consistently outperforms state-of-the-art augmentation strategies on various tasks, including ImageNet classification, weakly supervised object localization, and transfer learning for object detection and image captioning. 3. **Robustness and Uncertainty**: CutMix improves model robustness against input corruptions and reduces over-confidence in predictions. CutMix is implemented as a simple and efficient augmentation strategy, with minimal computational overhead. It has been shown to enhance the accuracy of CNNs on multiple datasets and architectures, demonstrating its effectiveness in various computer vision tasks. The source code and pre-trained models are available at <https://github.com/clovaai/CutMix-PyTorch>.CutMix is a novel data augmentation technique designed to enhance the performance of convolutional neural networks (CNNs) by improving both classification and localization capabilities. Unlike traditional regional dropout methods that remove informative pixels, CutMix replaces these regions with patches from other training images, maintaining the regularization effect while ensuring efficient use of training pixels. This approach allows the model to focus on less discriminative parts of objects, improving generalization and localization accuracy. The key contributions of CutMix include: 1. **Efficient Training**: By replacing removed regions with patches, CutMix avoids information loss and maintains training efficiency. 2. **Superior Performance**: CutMix consistently outperforms state-of-the-art augmentation strategies on various tasks, including ImageNet classification, weakly supervised object localization, and transfer learning for object detection and image captioning. 3. **Robustness and Uncertainty**: CutMix improves model robustness against input corruptions and reduces over-confidence in predictions. CutMix is implemented as a simple and efficient augmentation strategy, with minimal computational overhead. It has been shown to enhance the accuracy of CNNs on multiple datasets and architectures, demonstrating its effectiveness in various computer vision tasks. The source code and pre-trained models are available at <https://github.com/clovaai/CutMix-PyTorch>.