[Paper Review] PRIME: A few primitives can boost robustness to Common Corruptions

Apostolos Modas, Rahul Rade+3 more|arXiv (Cornell University)|Dec 23, 2021

Corruption and Economic Development5 citations

TL;DR

PRIME proposes a simple, principled data augmentation scheme using max-entropy families of basic image transformations—spectral (ω), spatial (τ), and color (γ)—to achieve state-of-the-art robustness against common image corruptions. By leveraging maximum entropy distributions with controlled smoothness (K) and strength (σ²), PRIME outperforms prior methods like AugMix and DeepAugment on ImageNet-C and CIFAR-10-C, even when used standalone or combined with other techniques.

ABSTRACT

This is the poster and presentation video for the paper "PRIME: A few primitives can boost robustness to Common Corruptions" presented at ECCV 2022.

Motivation & Objective

• Address the lack of principled, interpretable data augmentation methods for improving robustness to common image corruptions.
• Overcome the limitations of complex, heuristic-based augmentation pipelines that are hard to analyze and tune.
• Develop a general-purpose, plug-and-play data augmentation scheme that enhances robustness without sacrificing accuracy.
• Investigate the role of transformation diversity, mixing strategies, and online vs. offline augmentation in improving corruption robustness.

Proposed method

• Formalize a universal model of semantically preserving corruptions using additive subgroups of transformations in spectral, spatial, and color domains.
• Define transformation families using maximum entropy distributions under constraints of strength (σ²), smoothness (K), and boundary conditions.
• Implement three primitive transformations: ω (spectral, via random filters), τ (spatial, via displacement fields), γ (color, via random gain/offset).
• Construct PRIME augmentations by sampling from these max-entropy distributions and combining them via stochastic mixing (e.g., convex combinations).
• Apply the method in both online (per-batch) and offline (pre-augmented dataset) training schemes.
• Use analytical solutions derived from statistical physics to ensure tractable and unbiased sampling of transformations.

Experimental results

Research questions

• RQ1 Which fundamental design principles are essential for achieving robustness to common image corruptions?
• RQ2 How does the diversity of primitive transformations affect robustness, and is it more important than the complexity of the augmentation pipeline?
• RQ3 What is the role of mixing strategies in improving generalization across corrupted distributions?
• RQ4 How do online and offline data augmentation compare in terms of robustness and computational cost?
• RQ5 Can a simple, model-based approach outperform complex, end-to-end learned augmentation methods?

Key findings

• PRIME achieves state-of-the-art robustness on ImageNet-C, with a top-1 accuracy of 68.8% when used with online augmentation—surpassing AugMix (60.7%) and DeepAugment (67.5%) on the same setup.
• Even with only k=4 PRIME augmentations per image in offline training, PRIME achieves 88.8% accuracy on CIFAR-10-C, close to the 89.3% of online AugMix.
• The ablation study confirms that transformation diversity among primitives is crucial; methods lacking such diversity (e.g., AugMix) underperform PRIME.
• Mixing strategies play a pivotal role: PRIME’s performance drops significantly when mixing is removed, indicating its importance for generalization.
• Online augmentation with PRIME yields a +1.6% accuracy gain over offline training, with only 1.27× additional training time—making it highly efficient compared to complex methods like DeepAugment.
• PRIME improves robustness not only on standard benchmarks but also on perceptually dissimilar corruptions, artistic renditions, and adversarial filters, demonstrating broad generalization.