Improving robustness against common corruptions by covariate shift adaptation

TL;DR

This paper introduces an unsupervised online adaptation method that adjusts activation statistics in vision models to significantly improve robustness against common image corruptions, outperforming existing benchmarks.

Contribution

The authors propose a novel adaptation technique using unlabeled corrupted images to enhance model robustness, demonstrating substantial improvements across multiple models and benchmarks.

Findings

ResNet-50 improves from 76.7% to 62.2% mCE with adaptation.

State-of-the-art models like DeepAugment+AugMix achieve lower mCE scores after adaptation.

Even a single sample of corruption can enhance robustness.

Abstract

Today's state-of-the-art machine vision models are vulnerable to image corruptions like blurring or compression artefacts, limiting their performance in many real-world applications. We here argue that popular benchmarks to measure model robustness against common corruptions (like ImageNet-C) underestimate model robustness in many (but not all) application scenarios. The key insight is that in many scenarios, multiple unlabeled examples of the corruptions are available and can be used for unsupervised online adaptation. Replacing the activation statistics estimated by batch normalization on the training set with the statistics of the corrupted images consistently improves the robustness across 25 different popular computer vision models. Using the corrected statistics, ResNet-50 reaches 62.2% mCE on ImageNet-C compared to 76.7% without adaptation. With the more robust DeepAugment+AugMix model, we improve the state of the art achieved by a ResNet50 model up to date from 53.6% mCE to 45.4% mCE. Even adapting to a single sample improves robustness for the ResNet-50 and AugMix models, and 32 samples are sufficient to improve the current state of the art for a ResNet-50 architecture. We argue that results with adapted statistics should be included whenever reporting scores in corruption benchmarks and other out-of-distribution generalization settings.