PracticalDG: Perturbation Distillation on Vision-Language Models for Hybrid Domain Generalization

TL;DR

This work tackles Open Set Domain Generalization (OSDG) by distilling knowledge from large vision-language models (VLMs) into lightweight vision backbones using Perturbation Distillation (SCI-PD), which combines Score, Class, and Instance perturbations to robustly transfer semantics. It introduces Hybrid Domain Generalization (HDG), a benchmark that varies label-set overlap across source domains, along with the ${\rm H^{2}}$-CV metric to measure robustness across splits and data scarcity. The method avoids heavy fine-tuning of large models and demonstrates state-of-the-art performance on PACS, OfficeHome, and DomainNet while maintaining efficiency, with strong robustness under diverse HDG conditions. Overall, SCI-PD offers a practical, scalable solution for OSDG by leveraging CLIP-derived semantics to train compact models with improved generalization and reliability.

Abstract

Domain Generalization (DG) aims to resolve distribution shifts between source and target domains, and current DG methods are default to the setting that data from source and target domains share identical categories. Nevertheless, there exists unseen classes from target domains in practical scenarios. To address this issue, Open Set Domain Generalization (OSDG) has emerged and several methods have been exclusively proposed. However, most existing methods adopt complex architectures with slight improvement compared with DG methods. Recently, vision-language models (VLMs) have been introduced in DG following the fine-tuning paradigm, but consume huge training overhead with large vision models. Therefore, in this paper, we innovate to transfer knowledge from VLMs to lightweight vision models and improve the robustness by introducing Perturbation Distillation (PD) from three perspectives, including Score, Class and Instance (SCI), named SCI-PD. Moreover, previous methods are oriented by the benchmarks with identical and fixed splits, ignoring the divergence between source domains. These methods are revealed to suffer from sharp performance decay with our proposed new benchmark Hybrid Domain Generalization (HDG) and a novel metric $H^{2}$-CV, which construct various splits to comprehensively assess the robustness of algorithms. Extensive experiments demonstrate that our method outperforms state-of-the-art algorithms on multiple datasets, especially improving the robustness when confronting data scarcity.

Figures (6)

• Figure 1: The balance between model performance and training time consumption. Model performance is evaluated on the average H-score of different splits based on the proposed HDG benchmark. Our method achieves superior performance with less training time compared with state-of-the-art (SOTA) methods in OSDG.
• Figure 2: Illustration on the significance of the proposed HDG benchmark. Previous DG benchmarks are evaluated on a single split, producing unreliable conclusions for algorithms in practical usage. We claim that robust algorithms should possess stable performance on diverse data distributions.
• Figure 3: The overall framework of our method SCI-PD, including Score Perturbation (SP), Instance Perturbation (IP) and Class Perturbation (CP). SP saturates GT information into the similarity scores from CLIP to exploit semantics. IP excavates underlying semantics in instances via the weight distribution. CP saturates semantics from pretrained text embeddings to the class weights of the classifier.
• Figure 4: H-score on different domains under diverse hybridness $\mathcal{H}$ for OfficeHome.
• Figure 5: Experimental results on hyper-parameters.