sothatthedifferencecanbeseen. 5https: //pillow.readthedocs.io

InthecaseofCIFAR,allimages in the dataset are 32 by 32 pixels; we use the original images without anymodification at the test time. The original image was distorted by some corruptions, such as rotation and noise. As Table 5 shows, the two hypothetical loss predictors show significantly different performance. On the other hand, if the severity is high, the stronger blur effect (TSharpness:0.2) is used to drive performanceimprovement. Asintheseexamples,youcanseethat the test-time augmentation is determined according to the type of corruption or its severity.