Region Mixup

TL;DR

Region Mixup addresses the limitations of vanilla mixup by interpolating at the region level rather than across whole images. It tiles each image into a $k × k$ grid and forms a new sample by region wise fusion across $k^2$ regions using binary masks $M_j$ and interpolation weights $λ_j$, with $\sum M_j = 1$. On CIFAR-10, CIFAR-100 and Tiny ImageNet with PreAct ResNet-18, Region Mixup with $k=2$ achieves higher test accuracy than Vanilla Mixup and is competitive with or better than CutMix on several datasets. The paper also presents adversarial robustness tests and Grad-CAM++ visualizations. The approach is simple to implement with negligible overhead and broadens the regularization toolkit for visual recognition.

Abstract

This paper introduces a simple extension of mixup (Zhang et al., 2018) data augmentation to enhance generalization in visual recognition tasks. Unlike the vanilla mixup method, which blends entire images, our approach focuses on combining regions from multiple images.

Figures (2)

• Figure 1: Understanding region mixup
• Figure 2: Class activation mapping (CAM) Zhou2015LearningDF visualizations on Tiny ImageNet using Grad-CAM++ gradcam++.