On Interaction Between Augmentations and Corruptions in Natural Corruption Robustness

Invariance to a broad array of image corruptions, such as warping, noise, or\ncolor shifts, is an important aspect of building robust models in computer\nvision. Recently, several new data augmentations have been proposed that\nsignificantly improve performance on ImageNet-C, a benchmark of such\ncorruptions. However, there is still a lack of basic understanding on the\nrelationship between data augmentations and test-time corruptions. To this end,\nwe develop a feature space for image transforms, and then use a new measure in\nthis space between augmentations and corruptions called the Minimal Sample\nDistance to demonstrate a strong correlation between similarity and\nperformance. We then investigate recent data augmentations and observe a\nsignificant degradation in corruption robustness when the test-time corruptions\nare sampled to be perceptually dissimilar from ImageNet-C in this feature\nspace. Our results suggest that test error can be improved by training on\nperceptually similar augmentations, and data augmentations may not generalize\nwell beyond the existing benchmark. We hope our results and tools will allow\nfor more robust progress towards improving robustness to image corruptions. We\nprovide code at https://github.com/facebookresearch/augmentation-corruption.\n