Research map / Label-efficient and robust learning / Domain adaptation and generalization

Domain Generalization with Vital Phase Augmentation

AAAI 2024 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Optimization-based domain generalization · 12 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes VIPAug, a frequency-domain augmentation that splits 3D DFT phases into vital and non-vital sets by amplitude magnitude and perturbs them with different strengths (Gaussian noise and fractal phase replacement) to improve domain generalization on CIFAR and ImageNet corruptions.

Problem

Frequency-domain augmentation methods fix image phases and vary only amplitudes, which leaves models vulnerable to phase fluctuations under common corruptions and OOD shifts.

Approach

Identify vital phases (large-amplitude, more domain-invariant) via an argmax filter on the 3D DFT amplitude spectrum, apply weak Gaussian perturbations to vital phases and strong ones to non-vital phases, and replace non-vital phases with fractal phases, combined with amplitude replacement from APR-SP.

Key results

• On CIFAR-10, VIPAug reaches 95.8% clean accuracy and 8.4% corruption error, versus 95.3% and 25.3% for the baseline and 95.6% and 8.7% for APR-SP.
• On ImageNet-100, VIPAug + PRIME achieves the lowest mCE of 55.4%, while VIPAug alone gets 63.2% mCE versus 65.4% for APR-SP.
• On ImageNet, VIPAug reports 65.8% mCE, the best average among the compared methods, versus 67.4% for APR-SP and 68.4% for AugMix.

Measured effects reported in the paper

Surprising findings

• Reverse VIPAug (swapping vital and non-vital roles) performs worse than APR-SP without phase variation, suggesting phase variation does not always help.
• Using fractal images for VIPAug-F outperforms ImageNet, Stylized-ImageNet and GTA5 for phase replacement, which the authors attribute to classless images avoiding class mixing.

Limitations

• On CIFAR-100, the proposed method's clean accuracy (77.2-77.3%) is slightly below the baseline (78.3%), and PRIME lowers clean accuracy on CIFAR-10 and ImageNet-100.
• Hyperparameters such as filter size, variance and low-frequency retention ratio are dataset-specific and tuned empirically.
• 2D-DFT vs 3D-DFT comparison and ImageNet results are reported only for some settings; ImageNet baselines partly reuse numbers from AugMix rather than being re-run.

Experimental setup

Datasets: CIFAR-10 CIFAR-100 ImageNet-100 ImageNet Fractal images (DeviantArt)

Benchmarks: CIFAR-10-C CIFAR-100-C ImageNet-100-C ImageNet-C

Metrics: clean accuracy corruption error rate mean corruption error (mCE)

Baselines: Baseline APR-SP PRIME AugMix Patch Uniform AutoAugment Random AutoAugment MaxBlur pool SIN FACT HybridAugment++

Builds on: APR-SP PRIME AugMix AutoAugment

Closest papers by meaning

• Gradient-Guided Annealing for Domain Generalization (CVPR 2025 · similarity 0.84)
• Robust Domain Generalization under Divergent Marginal and Conditional Distributions (NeurIPS 2026 · similarity 0.84)
• AdvST: Revisiting Data Augmentations for Single Domain Generalization (AAAI 2024 · similarity 0.83)
• First-Order Manifold Data Augmentation for Regression Learning (ICML 2024 · similarity 0.83)
• PEER Pressure: Model-to-Model Regularization for Single Source Domain Generalization (CVPR 2025 · similarity 0.83)
• Sharpness-Aware Model-Agnostic Long-Tailed Domain Generalization (AAAI 2024 · similarity 0.83)