Pseudo Multi-Source Domain Generalization: Bridging the Gap Between Single and Multi-Source Domain Generalization

Shohei Enomoto

Abstract

Deep learning models often struggle to maintain performance when deployed on data distributions different from their training data, particularly in real-world applications where environmental conditions frequently change. While Multi-source Domain Generalization (MDG) has shown promise in addressing this challenge by leveraging multiple source domains during training, its practical application is limited by the significant costs and difficulties associated with creating multi-domain datasets. To address this limitation, we propose Pseudo Multi-source Domain Generalization (PMDG), a novel framework that enables the application of sophisticated MDG algorithms in more practical Single-source Domain Generalization (SDG) settings. PMDG generates multiple pseudo-domains from a single source domain through style transfer and data augmentation techniques, creating a synthetic multi-domain dataset that can be used with existing MDG algorithms. Through extensive experiments with PseudoDomainBed, our modified version of the DomainBed benchmark, we analyze the effectiveness of PMDG across multiple datasets and architectures. Our analysis reveals several key findings, including a positive correlation between MDG and PMDG performance and the potential of pseudo-domains to match or exceed actual multi-domain performance with sufficient data. These comprehensive empirical results provide valuable insights for future research in domain generalization. Our code is available at https://github.com/s-enmt/PseudoDomainBed.

Introduction

Deep learning models are notoriously brittle when they encounter data distributions different from their training set—a phenomenon known as "distribution shift." Domain Generalization (DG) aims to build models that can handle these shifts by learning features that are invariant across different environments. Currently, DG research is split into two primary paradigms: Multi-source Domain Generalization (MDG) and Single-source Domain Generalization (SDG).

While MDG is highly effective by leveraging diverse datasets (e.g., photos, cartoons, and sketches), it is often impractical due to the high cost and difficulty of multi-domain data collection. Conversely, SDG is more realistic for most applications but cannot directly utilize the sophisticated algorithms developed for the MDG setting. This paper addresses this "data gap" by asking: Can we use MDG's advanced logic even when we only have one source domain?

The PMDG Framework

To bridge this gap, we propose Pseudo Multi-source Domain Generalization (PMDG). The core idea is to treat a single source domain as a seed to grow a synthetic multi-domain dataset. This allows researchers to apply powerful MDG algorithms to the more practical SDG setting.

Overview of the PMDG framework. Transformations create pseudo-domains from a single source, which are then used to train a model using MDG algorithms.

The PMDG process follows three main steps:

  1. Pseudo-Domain Generation: Apply various transformations to the source data. We explored two main types: style transformations (e.g., AdaIN style transfer, CartoonGAN, and Edge Detection) and robust data augmentations (e.g., IPMix, AugMix, and RandConv).
  2. Domain Partitioning: Instead of pooling all augmented images into a single dataset, PMDG treats each transformation as a separate "pseudo-domain."
  3. MDG Training: These pseudo-domains are fed into standard MDG algorithms (such as CORAL, DANN, or SD) that are designed to align features across multiple sources.
Visualizations of various transformations used to generate pseudo-domains, including data augmentations like MixUp and CutMix, and style-based methods like CartoonGAN and Edge Detection.

Experimental Evaluation & Key Insights

We evaluated the framework on several standard DG benchmarks, including PACS, VLCS, OfficeHome, and TerraIncognita. We also introduced PseudoDomainBed, a modified version of the popular DomainBed benchmark specifically adapted for evaluating SDG and PMDG methods fairly.

Our analysis revealed several key insights:

  • Algorithm Transferability: There is a strong positive correlation between an algorithm's performance in a real multi-source setting and its performance in the PMDG setting. This suggests that the "logic" of MDG algorithms translates well to synthetic domains.
  • Efficiency of Synthetic Data: Surprisingly, under equal data volume conditions, models trained on pseudo-domains sometimes outperformed those trained on real multi-source data. This suggests that well-designed synthetic domains can be more "informative" for generalization than naturally occurring ones.
  • Architecture Sensitivity: The choice of transformation matters differently depending on the model backbone. For instance, style-based transformations that enhance shape features were more effective when using Vision Transformers (ViT) than ResNets.

Key Results

PMDG consistently outperformed standard SDG baselines across various architectures and datasets:

  • ResNet50: PMDG (using the 'SD' algorithm) achieved an average accuracy of 55.9%, surpassing the best SDG baseline, IPMix (55.2%).
  • Vision Transformer (ViT): PMDG reached 62.0% average accuracy, compared to 60.5% for the top baseline.
  • ImageNet: On large-scale tests, the PMDG approach showed better average accuracy across out-of-distribution variants (37.20%) than state-of-the-art SDG methods like IPMix (36.90%).

Conclusion

PMDG provides a robust framework for applying MDG logic to single-source scenarios. By leveraging pseudo-domains, we can achieve state-of-the-art generalization performance without the need for expensive multi-domain datasets. While limitations exist regarding the intensity and selection of transformations, these results provide a promising path forward for practical domain generalization research.

Summary updated

Abstract

Deep learning models often struggle to maintain performance when deployed on data distributions different from their training data, particularly in real-world applications where environmental conditions frequently change. While Multi-source Domain Generalization (MDG) has shown promise in addressing this challenge by leveraging multiple source domains during training, its practical application is limited by the significant costs and difficulties associated with creating multi-domain datasets. To address this limitation, we propose Pseudo Multi-source Domain Generalization (PMDG), a novel framework that enables the application of sophisticated MDG algorithms in more practical Single-source Domain Generalization (SDG) settings. PMDG generates multiple pseudo-domains from a single source domain through style transfer and data augmentation techniques, creating a synthetic multi-domain dataset that can be used with existing MDG algorithms. Through extensive experiments with PseudoDomainBed, our modified version of the DomainBed benchmark, we analyze the effectiveness of PMDG across multiple datasets and architectures. Our analysis reveals several key findings, including a positive correlation between MDG and PMDG performance and the potential of pseudo-domains to match or exceed actual multi-domain performance with sufficient data. These comprehensive empirical results provide valuable insights for future research in domain generalization. Our code is available at https://github.com/s-enmt/PseudoDomainBed.
BibTeX
@misc{enomoto2025pseudo,
  author = {Shohei Enomoto},
  title = {{Pseudo Multi-Source Domain Generalization: Bridging the Gap Between Single and Multi-Source Domain Generalization}},
  year = {2025},
  eprint = {2505.23173},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2505.23173},
}
Figure