Certainty and Uncertainty Guided Active Domain Adaptation

TL;DR

This work tackles Active Domain Adaptation by addressing a blind spot in prior work: the neglect of confident target predictions during sampling. It introduces a collaborative framework that combines Gaussian Process-based Active Sampling (GPAS) for uncertainty-driven target selection with Pseudo-Label-based Certain Sampling (PLCS) to inject high-confidence pseudo-labeled data, complemented by Uncertainty-balanced Class Sampling (UCS). The approach reduces the labeling search space while shrinking domain gaps, leading to substantial gains on Office-Home and DomainNet with a low annotation budget and competitive query efficiency. Extensive ablations demonstrate that each component—GPAS, PLCS, and UCS—contributes to improved adaptation, and the framework generalizes across ADA baselines, offering a practical path to efficient domain adaptation in real-world settings.

Abstract

Active Domain Adaptation (ADA) adapts models to target domains by selectively labeling a few target samples. Existing ADA methods prioritize uncertain samples but overlook confident ones, which often match ground-truth. We find that incorporating confident predictions into the labeled set before active sampling reduces the search space and improves adaptation. To address this, we propose a collaborative framework that labels uncertain samples while treating highly confident predictions as ground truth. Our method combines Gaussian Process-based Active Sampling (GPAS) for identifying uncertain samples and Pseudo-Label-based Certain Sampling (PLCS) for confident ones, progressively enhancing adaptation. PLCS refines the search space, and GPAS reduces the domain gap, boosting the proportion of confident samples. Extensive experiments on Office-Home and DomainNet show that our approach outperforms state-of-the-art ADA methods.

Figures (4)

• Figure 1: A) Previous Methods: These methods focus on selecting uncertain/informative samples for labeling but overlook confident samples, missing valuable information. B) Proposed Method: Our work aims to address this limitation by integrating confident pseudo-labels alongside uncertain samples, enhancing active domain adaptation by leveraging both to improve model accuracy.
• Figure 2: Our approach consists of two complementary phases per sampling round. GPAS ranks unlabeled target samples by posterior variance using class-wise GPs and queries labels for the top $b$ samples, reducing domain shift and increasing confident PL selection. PLCS selects the top $\kappa\%$ most confident target samples per class, adding them to the labeled set with their pseudo-labels. This, in turn, helps GPAS by shrinking the query search space.
• Figure 3: Comparison between different ADA approaches over AL rounds on three different domain transfers of Office-Home.
• Figure 4: Comparison between our method, SDM-AG, and entropy sampling with fixed $5\%$ active sampling rate while varying certain sampling rate from $0\%$ to $90\%$ on the Office-Home dataset.