Abstract
Conventional domain generalization aims to learn domain invariant representation from multiple domains, which requires accurate annotations. In realistic application scenarios, however, it is too cumbersome or even infeasible to collect and annotate the large mass of data. Yet, web data provides a free lunch to access a huge amount of unlabeled data with rich style information that can be harnessed to augment domain generalization ability. In this paper, we introduce a novel task, termed as semi-supervised domain generalization, to study how to interact the labeled and unlabeled domains, and establish two benchmarks including a web-crawled dataset, which poses a novel yet realistic challenge to push the limits of existing technologies. To tackle this task, a straightforward solution is to propagate the class information from the labeled to the unlabeled domains via pseudo labeling in conjunction with domain confusion training. Considering narrowing domain gap can improve the quality of pseudo labels and further advance domain invariant feature learning for generalization, we propose a cycle learning framework to encourage the positive feedback between label propagation and domain generalization, in favor of an evolving intermediate domain bridging the labeled and unlabeled domains in a curriculum learning manner.
Problem Definition: Semi-Supervised Domain Generalization
Domain Generalization (DG) aims to train models on several source domains so they generalize to an unseen target domain. Traditionally, this requires massive amounts of precisely labeled data across all source domains. However, in real-world scenarios, obtaining such labels is expensive and often impractical. While the web offers vast amounts of unlabeled data with diverse styles, standard Semi-Supervised Learning (SSL) fails because it assumes labeled and unlabeled data share the same distribution.
To address this, the authors introduce Semi-Supervised Domain Generalization (SSDG). This task involves learning from a mix of labeled and unlabeled source domains. The paper distinguishes between Close-set SSDG, where unlabeled data shares the same classes as labeled data, and Open-set SSDG, where unlabeled data (like web-crawled images) contains unknown categories, adding a layer of realistic difficulty.
Proposed Framework: Cyclic Learning with Evolving Intermediate Domains
The core challenge in SSDG is a fundamental conflict: pseudo-labeling works best when the domain gap is small, but generalization requires large domain gaps to learn style-invariant features. The proposed solution is a cyclic framework that progressively bridges this gap.
The framework consists of three main components:
- Adversarial Pseudo Labeling (APL): This module focuses on label propagation, using adversarial training to transfer class knowledge from labeled to unlabeled domains.
- Dual Calibrative Generalization (DCG): A dual-network architecture where models exchange "clean" samples identified via small-loss strategies. It includes a Style Confusor layer that shuffles feature statistics to force the model to focus on content rather than style.
- Evolving Intermediate Domain (EID): The central innovation. Instead of bridging the labeled and unlabeled domains in one step, the EID creates a "stepping stone" domain by blending labeled data with high-confidence unlabeled samples. This domain evolves over cycles, reducing the domain gap for APL and providing cleaner data for DCG in a curriculum learning manner.
Benchmarks and Experimental Results
The researchers evaluated their method on standard DG benchmarks (PACS, Digits-DG, Office-Home) adapted for SSDG. Crucially, they introduced two new web-crawled datasets to simulate real-world noise: PACS-Webdata and Office-Webdata. These datasets include significant portions of Out-of-Distribution (OOD) images, posing a realistic challenge for open-set SSDG.
The proposed framework consistently outperformed state-of-the-art Domain Adaptation (DA), DG, and SSL methods:
- PACS: Achieved 73.80% average accuracy, significantly higher than standard DA (~65%) and DG (~62%) baselines.
- Digits-DG: Reached 69.21% accuracy, compared to ~57% for traditional DA.
- Open-set Performance: On web-crawled benchmarks, the framework reached 75.67% on PACS-Webdata, demonstrating its ability to handle real-world noise and OOD samples effectively.
Conclusion
This work establishes SSDG as a vital research direction for practical robust vision. By decoupling label propagation from generalization and using an evolving intermediate domain, the proposed framework successfully harnesses unlabeled web data to improve model robustness. Future work will focus on better handling high ratios of OOD data in web-crawled sources and exploring more advanced domain synthesis techniques.
Summary updated