Center-aware Adversarial Augmentation for Single Domain Generalization

Tianle Chen, Mahsa Baktashmotlagh, Zijian Wang, Mathieu SalzmannWACV 2023 · OpenReview

Abstract

Single Domain Generalization (Single-DG) is a challenging task where a model must generalize to unseen domains after training on only one source environment. Traditional adversarial augmentation techniques often focus on generating "hard" examples near decision boundaries, which typically represent small, noisy perturbations rather than the large-scale distribution shifts seen in real-world domain changes. This paper introduces Center-aware Adversarial Domain Augmentation (CADA), a method that expands the source distribution by "pushing" samples away from their class centers in a latent space using a novel Angular Center Loss. By iterating between domain expansion and task optimization, CADA learns a more robust representation that generalizes significantly better to unseen styles and corruptions.

Introduction

Domain Generalization (DG) seeks to build models robust to shifts in data distribution, such as changes in lighting, background, or artistic style. While standard DG assumes access to multiple source domains, Single Domain Generalization (Single-DG) addresses the more restrictive and realistic scenario where only one source domain is available. Existing Single-DG methods frequently rely on adversarial training to create challenging examples. However, these "boundary-aware" methods often fail to simulate the substantial semantic-preserving shifts (e.g., photo to sketch) required for true generalization.

Methodology: The CADA Framework

The core of CADA is a min-max game driven by Angular Center Loss (AC), which measures the geodesic distance between features and their class centroids on a hypersphere.

The CADA framework: (Left) Initial state. (Center) Domain expansion pushes samples away from class centers. (Right) Task optimization pulls them back while maintaining class separation.

The process consists of two iterative steps:

  1. Domain Expansion (The "Push"): Pixels of source images are modified to maximize the Angular Center Loss. This pushes the latent features away from the class center in diverse directions, simulating a broader domain spread. A Mean Squared Error (MSE) constraint ensures the augmented images remain semantically consistent with the originals.
  2. Task Optimization (The "Pull"): The model is trained on these expanded samples by minimizing the Angular Center Loss. This pulls the diverse features back toward their respective class centers, effectively increasing the inter-class margin and creating a more compact, yet generalized, feature space.

Experimental Results

CADA was evaluated on several benchmarks, including Digits (MNIST to SVHN/USPS), CIFAR-10/100-C, and PACS.

  • Digits: CADA achieved an average accuracy of 80.56%, outperforming the previous state-of-the-art (74.56%).
  • CIFAR-10-C: The method showed exceptional robustness to corruptions like blur and noise, reaching 83.57% average accuracy compared to ~76% for competing methods.
  • PACS: CADA demonstrated scalability to complex real-world images, achieving a competitive average accuracy of 68.41%.

Qualitative Analysis and Localization

A key advantage of CADA is its ability to focus on essential object features rather than global style or background noise. Class Activation Maps (CAM) reveal that CADA-trained models exhibit superior localization.

Class Activation Maps (CAM) comparing the baseline ME-ADA with CADA across 15 types of image corruptions. CADA consistently focuses more accurately on the object of interest.

As shown in the CAM visualizations, while baseline methods like ME-ADA can be distracted by corruptions (e.g., fog, snow, or noise), CADA maintains a sharp focus on the primary subject. This suggests that the center-aware expansion forces the network to learn invariant features that define the class identity regardless of the domain shift.

Conclusion

CADA offers a novel perspective on Single Domain Generalization by shifting the focus from decision boundaries to class centers. By leveraging angular distance and an iterative expansion-optimization cycle, it effectively simulates large-scale domain shifts. While computationally more intensive than standard training due to the pixel-level optimization, its superior performance across diverse benchmarks highlights the potential of center-aware augmentation for building truly robust AI systems.

Summary updated

BibTeX
@inproceedings{chen2023center,
  author = {Tianle Chen and Mahsa Baktashmotlagh and Zijian Wang and Mathieu Salzmann},
  title = {{Center-aware Adversarial Augmentation for Single Domain Generalization}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2023},
  url = {https://openreview.net/forum?id=Wm0YCIaWHA},
}
Figure