[Paper Review] Unsupervised Domain Adaptation with Progressive Domain Augmentation

Kevin Hua, Yuhong Guo|arXiv (Cornell University)|Apr 3, 2020

Domain Adaptation and Few-Shot Learning40 references4 citations

TL;DR

This paper proposes Progressive Domain Augmentation (PrDA), a novel unsupervised domain adaptation method that generates virtual intermediate domains via Mixup-based data interpolation and performs progressive subspace alignment on the Grassmann manifold. By gradually bridging the domain shift through multiple alignment steps, PrDA achieves state-of-the-art performance on Office-31, Office-Caltech-10, and ImageCLEF-DA benchmarks, especially excelling in high-divergence domain adaptation scenarios.

ABSTRACT

Domain adaptation aims to exploit a label-rich source domain for learning classifiers in a different label-scarce target domain. It is particularly challenging when there are significant divergences between the two domains. In the paper, we propose a novel unsupervised domain adaptation method based on progressive domain augmentation. The proposed method generates virtual intermediate domains via domain interpolation, progressively augments the source domain and bridges the source-target domain divergence by conducting multiple subspace alignment on the Grassmann manifold. We conduct experiments on multiple domain adaptation tasks and the results shows the proposed method achieves the state-of-the-art performance.

Motivation & Objective

• To address the challenge of large domain shifts in unsupervised domain adaptation where direct alignment fails.
• To reduce information loss and improve alignment stability by decomposing large domain shifts into smaller, progressive steps.
• To enhance performance on standard benchmarks by leveraging intermediate virtual domains for smoother domain adaptation.
• To validate the effectiveness of progressive alignment in handling high inter-domain divergence.
• To demonstrate state-of-the-art performance across diverse domain adaptation tasks.

Proposed method

• The method generates a sequence of virtual intermediate domains through Mixup-based data interpolation between source and target domains.
• It performs multiple subspace alignments on the Grassmann manifold between the source domain and each virtual intermediate domain in a progressive manner.
• After each alignment step, pseudo-labeled data from the virtual domain is used to augment the source domain, gradually shifting its representation closer to the target.
• The process iteratively improves the alignment quality by reducing the domain divergence in stages rather than in a single step.
• The approach uses ResNet-50 as the feature extractor and relies on subspace alignment via projection matrices on the Grassmann manifold.
• It avoids direct adversarial training or reconstruction losses, focusing instead on structured, progressive domain alignment.

Experimental results

Research questions

• RQ1 Can progressive domain augmentation via virtual intermediate domains improve alignment stability and performance in unsupervised domain adaptation?
• RQ2 Does decomposing a large domain shift into multiple smaller alignment steps reduce information loss and improve generalization?
• RQ3 How does the proposed method perform on benchmarks with high inter-domain divergence, such as Office-Caltech-10?
• RQ4 Can progressive alignment outperform single-step domain alignment methods in terms of accuracy and robustness?
• RQ5 How does the method compare to state-of-the-art approaches like Ad-REM, MEDA, and CDAN across diverse transfer tasks?

Key findings

• PrDA achieved state-of-the-art performance on 7 out of 12 transfer tasks in the Office-Caltech-10 benchmark.
• On the Office-Caltech-10 dataset, PrDA achieved 99.7% accuracy on the D→W task and 99.3% on the A→W task, outperforming all baselines.
• In the ImageCLEF-DA benchmark, PrDA achieved 81.8% accuracy on I→P, 91.0% on P→I, and 95.5% on P→C, outperforming DANN, CDAN, and MEDA on several tasks.
• The method demonstrated strong performance on high-divergence tasks, with domain classifier accuracy reaching 76.1% on A→D and D→A, indicating significant domain shift.
• PrDA achieved competitive results even on low-divergence tasks, showing robustness across varying domain similarity levels.
• The ablation study confirmed that progressive alignment with virtual domains significantly improves performance compared to direct alignment.