Research map / Label-efficient and robust learning / Test-time adaptation
On Interaction Between Augmentations and Corruptions in Natural Corruption Robustness
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train with augmentations perceptually similar to a test corruption (lower MSD) | WideResNet-40-2 on CIFAR-10 augmentation powerset · CIFAR-10-C | corruption error: 12 of 15 corruptions have Spearman correlation > 0.6 | Helped |
| use MMD instead of MSD as the augmentation-corruption distance (ablation) | augmentation powerset · CIFAR-10-C | correlation with corruption error | Hurt |
| use VGG-19-BN instead of WideResNet as MSD feature extractor (ablation) | MSD · CIFAR-10-C | Spearman correlation: 0.90 -> 0.92 (Impulse Noise) | No clear effect |
| evaluate on dissimilar corruptions (C-bar) instead of ImageNet-C | Patch Gaussian ResNet-50 · ImageNet | error (%): 49.3 -> 56.6 | Hurt |
| evaluate on dissimilar corruptions (C-bar) instead of ImageNet-C | AutoAugment ResNet-50 · ImageNet | error (%): 55.0 -> 55.7 | Hurt |
| evaluate on dissimilar corruptions (C-bar) instead of CIFAR-10-C | AugMix WideResNet-40-2 · CIFAR-10 | error (%): 11.1 -> 16.0 | Hurt |
| evaluate non-augmentation robust models on ImageNet-C-bar | WSL, EfficientNet, Noisy Student, ViT-S, ResNeSt · ImageNet | error (%): ViT-S 44.5 -> 41.1; WSL 38.1 -> 39.2 | Mixed |
| sample only the most similar AugMix augmentations (about 100) instead of full AugMix (ablation) | AugMix on CIFAR-10-C · CIFAR-10-C | corruption error: about 1% gain | Helped |
| sample only the most dissimilar AugMix augmentations (ablation) | AugMix on CIFAR-10-C · CIFAR-10-C | corruption error: around 5% drop in accuracy | Hurt |
| use only about 1000 random AugMix transforms instead of all sampled ones (ablation) | AugMix on CIFAR-10-C · CIFAR-10-C | corruption error | No clear effect |
Closest papers by meaning
- Adaptive Sensitivity Analysis for Robust Augmentation against Natural Corruptions in Image Segmentation (ICML 2025 · similarity 0.88)
- Improving robustness to corruptions with multiplicative weight perturbations (NeurIPS 2024 · similarity 0.86)
- HYBOOD: A Hybrid Generative Model for Out-of-Distribution Detection with Corruption Estimation (AAAI 2025 · similarity 0.85)
- Slight Corruption in Pre-training Data Makes Better Diffusion Models (NeurIPS 2024 · similarity 0.85)
- Fourier-basis Functions to Bridge Augmentation Gap: Rethinking Frequency Augmentation in Image Classification (CVPR 2024 · similarity 0.84)
- Mint: A Simple Test-Time Adaptation of Vision-Language Models against Common Corruptions (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.