SUMix: Mixup with Semantic and Uncertain Information

TL;DR

A novel approach named SUMix is proposed to learn the mixing ratio as well as the uncertainty for the mixed samples during the training process, and a learnable similarity function is designed to compute an accurate mix ratio.

Abstract

Mixup data augmentation approaches have been applied for various tasks of deep learning to improve the generalization ability of deep neural networks. Some existing approaches CutMix, SaliencyMix, etc. randomly replace a patch in one image with patches from another to generate the mixed image. Similarly, the corresponding labels are linearly combined by a fixed ratio $λ$ by l. The objects in two images may be overlapped during the mixing process, so some semantic information is corrupted in the mixed samples. In this case, the mixed image does not match the mixed label information. Besides, such a label may mislead the deep learning model training, which results in poor performance. To solve this problem, we proposed a novel approach named SUMix to learn the mixing ratio as well as the uncertainty for the mixed samples during the training process. First, we design a learnable similarity function to compute an accurate mix ratio. Second, an approach is investigated as a regularized term to model the uncertainty of the mixed samples. We conduct experiments on five image benchmarks, and extensive experimental results imply that our method is capable of improving the performance of classifiers with different cutting-based mixup approaches. The source code is available at https://github.com/JinXins/SUMix.

Figures (10)

• Figure 1: The diagram shows different cases of raw samples that underwent the CutMix with a mixing ratio of 0.5 to obtain mixed samples. $Left:$ the raw samples and their one-hot labels; $Right:$ labels of the mixes obtained using a redefined mixing ratio $\widetilde{\lambda}$ for different cases.
• Figure 2: The figure shows some hand-crafted mixup methods with "Label MisMatch" problem. In CutMix and ResizeMix, the "Shark" is mostly obscured, while in SaliencyMix and FMix, it is difficult to notice the "Panda".
• Figure 3: The results illustrate the application of SUMix. Left: Top-1 Accuracy improvement from the mixup approaches with SUMix; Right: Comparison of Vanilla method, CutMix, and SUMix for CAM visualization.
• Figure 4: The left diagram represents the training process of SUMix. The raw and mixed samples are encoded through the model to obtain the semantic and uncertainty information, and the new mix ratio and loss function for regularization are obtained. The right diagram is the spatial distance between the raw and mixed samples.
• Figure 5: The top of the figure shows a visualization of the sample at 0% to 100% occlusion ratio. The light blue curves in the lower four subfigures show the classification accuracy of CutMix, FMix, SaliencyMix, and ResizeMix on ImageNet-1K using ResNet18 for 100 epochs of training, respectively, and the red curve shows the classification accuracy of these methods using SUMix.