[Paper Review] Agnostic Domain Generalization.

Fabio Maria Carlucci, Paolo Russo+2 more|arXiv (Cornell University)|Aug 3, 2018

Domain Adaptation and Few-Shot Learning36 references12 citations

TL;DR

This paper proposes a domain generalization method that hallucinates agnostic images from multiple source domains to improve cross-domain generalization without requiring labeled or even unlabeled target data. By using a deep architecture with dual adversarial domain classifiers at image and feature levels, the model disentangles domain-specific style from categorical knowledge, achieving state-of-the-art performance in both multi-source domain adaptation and domain generalization settings.

ABSTRACT

The ability to generalize across visual domains is crucial for the robustness of artificial recognition systems. Although many training sources may be available in real contexts, the access to even unlabeled target samples cannot be taken for granted, which makes standard unsupervised domain adaptation methods inapplicable in the wild. In this work we investigate how to exploit multiple sources by hallucinating a deep visual domain composed of images, possibly unrealistic, able to maintain categorical knowledge while discarding specific source styles. The produced agnostic images are the result of a deep architecture that applies pixel adaptation on the original source data guided by two adversarial domain classifier branches at image and feature level. Our approach is conceived to learn only from source data, but it seamlessly extends to the use of unlabeled target samples. Remarkable results for both multi-source domain adaptation and domain generalization support the power of hallucinating agnostic images in this framework.

Motivation & Objective

• To address the challenge of domain generalization when unlabeled target data are unavailable in real-world settings.
• To develop a method that learns robust representations from multiple source domains without relying on target domain data.
• To explore whether hallucinated, agnostic images—realistic in content but style-agnostic—can improve generalization across visual domains.
• To design a training framework that seamlessly incorporates unlabeled target samples when available, enhancing flexibility.

Proposed method

• The method employs a deep generative architecture that hallucinates images from source data, preserving semantic content while discarding domain-specific styles.
• It uses two adversarial domain classifiers: one operating at the image level and another at the feature level, both trained to disentangle domain-specific characteristics.
• The image-level discriminator ensures that generated images are visually plausible but style-agnostic, while the feature-level discriminator enforces domain-invariant representations.
• The model is trained solely on source data, yet it generalizes effectively to unseen domains, even when unlabeled target data are later introduced.
• The framework enables end-to-end learning of domain-invariant features through adversarial training and pixel-level adaptation.
• The approach is extensible to include unlabeled target samples during training, improving performance without requiring annotations.

Experimental results

Research questions

• RQ1 Can a model generalize effectively across visual domains when no target domain data are available during training?
• RQ2 How effective is image hallucination in preserving semantic content while removing domain-specific styles for improved generalization?
• RQ3 To what extent can a model trained only on source data achieve performance comparable to methods using target data?
• RQ4 Can the proposed method be extended to include unlabeled target samples without retraining?
• RQ5 Does dual adversarial training at image and feature levels lead to better domain-invariant representations than single-level approaches?

Key findings

• The proposed method achieves state-of-the-art performance on multi-source domain adaptation benchmarks, even without using any target data during training.
• The hallucinated agnostic images effectively preserve categorical knowledge while discarding source-specific styles, enabling strong generalization.
• The model generalizes well to unseen domains, demonstrating robustness in zero-shot domain generalization settings.
• When unlabeled target data are incorporated, performance improves further, showing the method's flexibility and scalability.
• The dual adversarial training at image and feature levels leads to more disentangled and domain-invariant representations than single-level approaches.
• The approach outperforms existing domain generalization and unsupervised domain adaptation methods on standard benchmarks, confirming its effectiveness.