PaperlayerView exact revision on arXiv

Source-linked AI summary

Improving Category-Aware Active Domain Adaptation via Uncertainty-Guided Region Coverage

My Tra Nguyen Huynh, Viet Cuong Ta

arXiv:2610.19803v1cs.CV

TL;DR

Average-accuracy evaluation in active domain adaptation can leave difficult categories behind, while influence-based point selection can be sensitive and redundant. The paper proposes uncertainty-guided region selection using positive-manifold coverage, and reports the highest average accuracy per category across all tasks of three benchmarks, including 87.73% on Office-Home.

  • Problem

    Average-accuracy evaluation in active domain adaptation can favor easier categories, while influence-based point selection may be parameter-sensitive and redundant.

    Source: §1 · p. 2

  • Method

    The method selects regions using Monte-Carlo positive-manifold coverage and uses uncertainty guidance to focus coverage toward decision-boundary regions.

    Source: §1 · p. 2, §3.2 · p. 9

  • Results

    The method achieved the highest average accuracy per category in all tasks across three benchmarks, reaching 87.73% on Office-Home.

    Source: §4.2 · p. 14

  • Takeaways & Limitations

    The approach is designed to improve individual-category performance while representing each category’s target distribution.

    Source: §5 · p. 18

    The conclusion identifies finer sampling as a possible way to balance the scoring mechanism’s uncertainty-coverage trade-off.

    Source: §5 · p. 19

Abstract

from arXiv · show

Active domain adaptation researches focus on alleviating the annotation cost, which comprises source data and target data labelling, as well as enhancing the performance of unsupervised domain adaptation. However, optimizing the adaptation model based on the average accuracy can lead to the drop in performance of some difficult categories. To overcome this issue, a point-wise selection strategy built upon the influence function can be used to improve domain adaptation performance. In our work, we first address the sensitive and redundant aspects of point-wise selection, which can lead to overfitting of the acquisition function. We thus introduce a region-based approach that ensures the improvement of the adaptation model by maximizing a positive manifold coverage. An uncertainty score is further introduced to the selection mechanism to learn a better decision boundary for the target domain. In comprehensive experiments, our proposed method surpasses the first category-aware active domain adaptation baseline as well as existing active domain adaptation methods across multiple domain adaptation benchmarks. Our implementation is available at: https://github.com/nhtm2702/WHO IS NEXT.

1 Introduction

Active domain adaptation can reduce labeling costs, but average-accuracy selection may leave difficult categories behind. The paper motivates category-aware region selection with uncertainty guidance to address point-selection sensitivity and redundancy.

1 Introduction

  • Average-performance evaluation can favor easier categories, while challenging categories receive little improvement or may lose performance after active-learning rounds.The introduction frames this as a bias of existing active domain adaptation methods.Source: §1 · p. 2
  • CAADA uses influence-based point selection for individual categories, but early model unreliability and locally concentrated selections can make scores parameter-sensitive and redundant.Close queried samples may provide overlapping information and risk overfitting the acquisition function to the empirical objective.Source: §1 · p. 2
  • The proposed method replaces point selection with region selection using Monte-Carlo positive-manifold coverage, then uses uncertainty guidance to focus coverage near decision boundaries.The introduction reports evaluation across three benchmarks, with significantly higher average accuracy and better underlying data coverage.Source: §1 · p. 2

2 Background and Preliminaries

The background situates active domain adaptation as cost-conscious target labeling and outlines both prior selection strategies and the paper’s category-aware setting. It then describes influence-based selection and kernel herding as foundations for the method.

2.1 Related Works

  • Active learning selects a subset of unlabeled data for annotation; prior active domain adaptation methods combine uncertainty, diversity, or evidential uncertainty to guide that selection.The related-work discussion names CLUE and DUC [4] as examples.Source: §2.1 · p. 3

2.2 Background

  • Active domain adaptation annotates the full source dataset and a selected target subset, iteratively querying unlabeled target samples and updating the labeled pool.The setting assumes the source and target domains share the same category set.Source: §2.2 · p. 3
  • For each category, the background describes a binary logistic-regression classifier and an influence predictor trained to estimate labeled-sample influence using a Hessian-based calculation.The highest-scoring B unlabeled target samples are selected, and category-classifier accuracy is reported after K rounds.Source: §2.2 · pp. 3–4
  • Kernel herding provides a basis for modeling target-distribution coverage through an embedding-space similarity measure, where closer samples receive higher similarity.The passage notes that this measurement implicitly depends on a good embedding space correlated with semantic similarity.Source: §2.2 · p. 4

3 Method

Section 3 presents a category-aware selection method that uses uncertainty-guided positive-manifold region coverage to reduce redundancy in target-region selection. Its analysis explains local concentration under influence-based selection and gives conditions under which uncertainty weighting preserves coverage gains up to multiplicative constants.

3.1 Motivation

  • Influence-based selection depends on the fitted regression model and can be sensitive to parameter changes; selected points may also cluster in dense local regions.The paper contrasts this with positive-manifold coverage, which restricts selection to positive, category-relevant regions, unlike whole-distribution coverage that may include irrelevant regions.Source: §3.1 · pp. 4–5

3.2 Uncertainty-Guided Selection with Maximizing Region Coverage

  • Positive-manifold coverage targets category-specific regions and combines coverage gain with uncertainty to prioritize informative regions near the decision boundary.Uncertainty guides the coverage estimate rather than directly selecting a sample; the score accounts for uncertainty in both the candidate and the region it covers.Source: §3.2 · pp. 5, 6, 8, 9

3.3 Theory Analysis

  • Under a Lipschitz assumption, high influence can concentrate top-B selections near a high-influence sample, while region-based coverage favors candidates that reduce redundancy.The analysis also states that bounded uncertainty weights preserve coverage gain, assuming the uncertainty signal is bounded below by α > 0.Source: §3.3 · pp. 9–12
  • Uncertainty weighting preserves the positive region coverage gain up to multiplicative constants.Source: §3.3 · p. 12
  • The method combines CAADA’s category awareness with MaxHerding’s region coverage [18], restricting coverage to the target class’s positive region and guiding selection with uncertainty.Unlike class-agnostic coverage, it favors candidates with greater positive-region coverage gain when category relevance is comparable, reducing redundancy from independent top-B pointwise selection; uncertainty weighting is supported as more effective than either signal alone in active domain adaptation.Source: §3.3 · p. 13

4 Experiments

Experiments compare the method with active domain adaptation baselines across Office-Home, DomainNet-126, and VisDA-2017. The method achieves the highest average per-category accuracy across the benchmarks, while analyses examine sample coverage, component contributions, and scoring choices.

4.1 Settings

  • The experiments use Office-Home, DomainNet-126, and VisDA-2017, comparing against five active domain adaptation baselines and CAADA under a shared protocol.Office-Home uses a target-data budget of 1%; DomainNet-126 and VisDA-2017 use a budget of 100 samples.Source: §4.1 · p. 13

4.2 Main results

  • The method achieves the highest average per-category accuracy across all tasks of three benchmarks, averaging 87.73% on Office-Home, 83.93% on DomainNet-126, and 79.37% on VisDA-2017.It outperforms CAADA by an average 8.41% on Office-Home, 26.16% on DomainNet-126, and 12.69% on Synthetic-to-Real VisDA-2017.Source: §4.2 · p. 14, Table 1 · p. 14, Table 2 · p. 14

4.3.1 Analysis on the quality of selection samples

  • On Batteries and Trash Can, the method covers category distributions more broadly than CAADA, whose selections concentrate near the center and overlook distribution tails.For Curtains, its more uniform target-manifold sampling has lower MMD, a 4.76% performance gain, and 97.62% accuracy; with budget 10, MMD stabilizes after about 10 rounds while accuracy continues rising toward convergence.Source: §4.3.1 · pp. 15–16, Figure 4 · p. 15, Figure 5 · p. 16

4.3.2 Analysis on contribution of each component

  • Coverage Only improves performance over CAADA by 6.61%, and adding weighted regions raises the improvement to 14.15%, with about 0.05% standard deviation.Together, the modules improve performance by 20.76%; Coverage Only and the full method reduce runtime by about 94.20% and 87.94%, respectively, versus CAADA. The selected samples also have higher uncertainty than those selected by Coverage Only.Source: §4.3.2 · pp. 16–18, Table 3 · p. 17, Figure 8 · p. 17

4.3.3 Analysis on the scoring mechanism

  • Changing the negative-manifold weight varies accuracy by only about 0.07%, with no consistent improvement, so the main experiments use only the positive manifold.The comparison concerns combinations of positive and negative manifolds on the Rw→Pr Office-Home task.Source: §4.3.3 · p. 18, Table 4 · p. 18

5 Conclusion

The paper introduces category-aware active domain adaptation using region-based positive manifold coverage and uncertainty guidance. It reports improvements across benchmarks and identifies joint fine-tuning and finer sampling as possible extensions.

5 Conclusion

  • The method outperforms category-aware and other active domain adaptation methods across various benchmarks, with significant performance improvements.The conclusion does not specify numerical results.Source: §5 · p. 18
  • Region-based positive manifold coverage replaces point-wise influence-based selection to reduce redundancy and parameter sensitivity while expanding represented target-class regions.Coverage gain is approximated through Monte-Carlo estimation.Source: §5 · p. 18
  • Theoretical analysis explains influence-only selection's local concentration and establishes redundancy reduction for the region-based objective.Uncertainty weighting preserves positive coverage gain up to a multiplicative constant.Source: §5 · p. 18
  • Joint fine-tuning of the extractor and classifier, and finer sampling to balance uncertainty and coverage, are proposed as extensions.Source: §5 · p. 19
Loading 2610.19803v1…