Pith. sign in

REVIEW 4 major objections 3 minor 71 references

Easy-first two-stage training sets new source-free adaptation records

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 11:41 UTC pith:FM5S3E3W

load-bearing objection A coherent SFDA combination with a real formal bug in the central graph-contrastive loss and test-set-tuned hyperparameters; worth refereeing after fixes. the 4 major comments →

arxiv 2607.29202 v1 pith:FM5S3E3W submitted 2026-07-31 cs.CV

Domain-Division based Progressive Learning for Source-Free Domain Adaptation

Pan Liu , Jing Li , Meng Zhao , Wanli Xue , Qinghua Hu , Shengyong Chen This is my paper
classification cs.CV
keywords source-free domain adaptationdomain divisionprogressive learninguncertainty-aware self-traininggraph contrastive learninginstance contrastive learningpseudo labelingconfidence and entropy
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that a source-pretrained model can be adapted to an unlabeled target domain without any source data by first splitting the target into easy-to-adapt and hard-to-adapt subdomains according to prediction confidence and uncertainty, then training in two alternating stages that begin with the easy samples and progress to the hard ones. The easy/hard division is recomputed at the start of each stage, so the curriculum tracks the model's improving predictions. Stage one uses uncertainty-weighted self-training and a graph contrastive loss to pull hard samples toward reliable easy samples of the same pseudo class; stage two applies consistency learning to easy samples and instance contrastive learning within the hard subdomain, exploiting the target's local structure without trusting hard-sample pseudo labels. On three standard benchmarks the method claims state-of-the-art average accuracies, including on VisDA, where it also outperforms adaptation methods that still have access to source data. A sympathetic reader would care because the core idea—progressive learning from easy to hard within the target domain—offers a portable way to use samples that most self-training methods discard.

Core claim

The paper's central claim is that per-sample adaptation difficulty in the target domain is readable from the model's own predictions, and that organizing training around this difficulty—easy samples first, then hard samples, with separate strategies for each—is enough to match or exceed the accuracy of source-free methods that rely on more elaborate filtering or generation. Concretely, samples with high predicted confidence and low predictive entropy are collected class-by-class into an 'easy' subdomain; everything else is 'hard'. Two alternating stages alternately use uncertainty-aware self-training with graph contrastive alignment across subdomains, then targeted learning on each subdomain

What carries the argument

The central object is the domain-division operation: at the start of each stage the target set is partitioned class-wise into an easy subdomain (top confidence per class, then lowest entropy within that candidate set) and a hard subdomain (the rest). This split drives everything that follows: pseudo labels are initially computed from easy-sample cluster centers, then from neighborhood soft voting; stage-one losses (entropy-weighted cross-entropy, graph contrastive loss between a label graph and a feature graph, and nuclear-norm maximization) are designed to align hard samples with easy ones; stage-two losses (standard self-training and augmentation-consistency on easy samples, instance contr

Load-bearing premise

The division into easy and hard subdomains assumes that high-confidence, low-entropy predictions in the target domain are predominantly correct classifications; if domain shift makes the model confidently wrong, the easy set is poisoned and the uncertainty-aware, graph-alignment, and consistency losses all compound the error — the paper's own VisDA car-category note is an admission of this failure mode.

What would settle it

Measure the label accuracy of the easy subdomain at the first epoch on a target with strong domain shift; if E-acc is not substantially above the accuracy on all target samples, or if a deliberately inverted split (selecting high-entropy samples as 'easy') performs equally well or better, the progressive-learning premise collapses. The paper's E-acc plots, which show easy accuracy well above all-sample accuracy, are the current evidence; an independent implementation on a miscalibrated model—confidently wrong on a large fraction—should be run to test it.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the split is reliable, samples that most pseudo-label pipelines discard as noisy can be actively exploited through instance contrastive learning, raising the ceiling for source-free adaptation.
  • Recomputing the division each stage lets the curriculum deepen as the model improves, so the method does not depend on a one-time clean/dirty split.
  • The reported averages on Office-31, Office-Home, and VisDA imply the method is competitive with—and on VisDA better than—adaptation methods that still access source data, suggesting source-free does not have to cost accuracy.
  • The paper's ablation ties each loss to a specific function: graph contrastive alignment readies hard samples, uncertainty weighting stabilizes early training, and the two-stage refinement yields the final gain.
  • The near-zero nuclear-norm coefficients on larger datasets indicate the regularization's benefit shrinks as target size grows, a scaling clue for future pipelines.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A testable extension is to feed the easy/hard split into a black-box adaptation setting (only classifier outputs, no feature access) — the paper's softmax-space losses are directly portable there, even though the current experiments use full model access.
  • The method's sensitivity to class imbalance is hinted by the paper's own note on VisDA's car category; a class-balanced or adaptive-neighbor variant is a natural and falsifiable follow-up.
  • The load-bearing premise—that confidence and entropy identify correctness in the target domain—is validated only through E-acc curves; perturbing the division (e.g., random or deliberately wrong splits) would be a sharper causal test of the mechanism.
  • Because gamma (easy-set fraction) and T (neighbor count) are tuned per dataset, the paper does not establish a universal rule for these constants; a calibration procedure that sets them from target statistics would strengthen the approach beyond the three benchmarks.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 3 minor

Summary. The paper proposes DPL, a source-free domain adaptation method that first divides the unlabeled target domain into easy-to-adapt and hard-to-adapt subsets using class-wise confidence and entropy, then alternates two training stages. Stage one performs uncertainty-weighted self-training with pseudo labels and aligns the two subdomains through graph contrastive learning. Stage two applies consistency self-training to the easy subset and instance contrastive learning using local neighborhood structure to the hard subset, with nuclear-norm regularization in both stages. The method is evaluated on Office-31, Office-Home, and VisDA against UDA and SFDA baselines, reporting average accuracies of 90.7%, 74.0%, and 87.8%, respectively. The paper includes ablations for each loss component, hyperparameter sensitivity analyses, and makes code available.

Significance. If the reported results are reliable, DPL would be a useful contribution to SFDA: it explicitly addresses the under-exploitation of hard samples in self-training, has a clean two-stage progressive structure, and ships with code. The ablation study in Table IV is a genuine strength, as it isolates the contribution of each loss term, and the E-acc analysis in Fig. 5 gives some evidence for the central easy/hard split assumption. However, the formal problem in the graph-contrastive objective and the test-set-based hyperparameter selection currently prevent the numeric SOTA claims from being accepted at face value.

major comments (4)
  1. [§III-D2, Eq. (14)-(15)] The graph-contrastive loss is mathematically undefined as written. Eq. (14) defines A_ij = q_i · q_j with normalized features, so A_ij can be negative (or zero). Eq. (15) then minimizes -1/N_b Σ_i Σ_j R_ij log A_ij, which requires log of non-positive values. The sentence 'both graphs are row-wise normalized' does not repair this: row-sum normalization of a signed matrix preserves negative entries, and no softmax/exp normalization is stated. This is load-bearing because L_gcl is a central stage-one objective and Table IV credits the GCL+ICL combination with 0.6/0.5 point gains. Please replace Eq. (14) with a positive affinity (e.g., exp(q_i·q_j/τ) followed by row normalization) and confirm that the reported experiments use that corrected objective.
  2. [§IV-B (hyperparameters)] Hyperparameters γ, T, α, β, and λ are chosen per dataset, and the text says 'For λ, we use the value with the best performance.' The sensitivity plots in Fig. 3–4 are computed on the same test benchmarks. This is effectively test-set tuning. Since the SOTA margins are small (Office-31 Avg. 90.7 ties MLSL and is only +0.5 over TPDS; Office-Home +0.5 over TPDS; VisDA +0.2 over TPDS), the comparison is not yet convincing. Report a validation-based hyperparameter selection, fix hyperparameters across datasets, or demonstrate that the ordering is stable under a single setting.
  3. [Tables I–IV] All accuracy numbers are point estimates without error bars or multiple-seed results. For claims of 'outperforming all compared methods' that rest on average improvements below one percentage point (e.g., Office-Home 74.0 vs 73.5 for TPDS; VisDA 87.8 vs 87.6 for TPDS), single-run numbers are insufficient. Please report mean and standard deviation over at least three runs, and state the number of seeds used.
  4. [§IV-E, Fig. 5] The progressive training scheme rests on the premise that the confidence/entropy division identifies samples that are not only easy for the model but also correctly classified. The only direct evidence is the E-acc curves in Fig. 5, which cover four Office-Home tasks. Because a confidently wrong easy set would be amplified by self-training, graph alignment, and consistency losses, please provide E-acc (or class-wise purity) curves for all three benchmarks, or at least for representative tasks from each.
minor comments (3)
  1. [§III-C, Eq. (8)] The 'T neighboring samples' used for soft-voting pseudo labels are not defined. Specify the neighbor search space (e.g., cosine distance in the feature space using the memory bank) and whether T is the same as in stage two.
  2. [§II-B, references] The sentence 'BETA relies on two-network co-training and the MixMatch semi-supervised learning approach [37]' cites [37], which is DivideMix, not MixMatch. Please correct the citation or the reference.
  3. [§III-E1, Eqs. (16)-(17)] The notation p_i vs p'_i is introduced only in prose; define in the equations that p_i is the weak-augmentation softmax output and p'_i is the strong-augmentation output before using them in L_ce2 and L_cs.

Circularity Check

0 steps flagged

No circularity: the central claim is benchmark-verified and the design borrowings are external; the only self-citation is peripheral.

full rationale

The paper's derivation chain is a proposed training algorithm evaluated on standard SFDA benchmarks. The easy/hard domain division uses confidence and entropy as an explicit operational proxy, and its validity is checked empirically via E-acc curves rather than asserted by definition. Pseudo-label generation, uncertainty weighting, graph contrastive alignment, and instance contrastive learning are all specified with explicit equations and are trained against current-model outputs; final accuracy is measured on held-out target labels, not recovered from fitted values. The graph-contrastive loss is attributed to SEAL [36], which is an external work, and the only author-overlapping citation [48] is used for a peripheral design choice (output-space contrastive learning), not as a load-bearing uniqueness or derivation step. There is no fitted parameter that is later renamed as a prediction, and no self-citation chain is invoked to force the method's choice. The formal issue in Eq. (15) — taking the logarithm of raw cosine similarities A_ij = q_i·q_j, which can be non-positive — is a real correctness/reproducibility concern, but it is not circularity: it does not make the claimed result equivalent to its inputs by construction. Similarly, the reported choice of hyperparameters "with the best performance" is an evaluation-hygiene matter, not a derived-then-predicted circular step. The paper is therefore self-contained with respect to circularity, and no circular step is identified.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

DPL introduces no new physical or conceptual entities. Its central claim rests on standard domain-adaptation assumptions, a set of dataset-specific hyperparameters selected with test-set performance, and an underspecified normalization step in the graph contrastive loss.

free parameters (4)
  • gamma (easy-set selection ratio) = 0.4 (Office-31), 0.5 (Office-Home), 0.5 (VisDA)
    Controls the fraction of high-confidence low-entropy samples selected per class in domain division; chosen per dataset.
  • T (number of neighbors) = 5 (Office-31), 4 (Office-Home), 5 (VisDA)
    Neighborhood size used for soft-voting pseudo labels and instance contrastive positives; tuned per dataset.
  • lambda (graph contrastive loss weight) = 0.7 (Office-31), 0.6 (Office-Home), 0.6 (VisDA)
    The paper states 'For lambda, we use the value with the best performance', meaning it was selected using test-set accuracy.
  • alpha and beta (nuclear-norm loss weights) = alpha=3/beta=2 (Office-31), alpha=0/beta=3 (Office-Home), alpha=0/beta=0 (VisDA)
    Chosen per dataset after observing the heatmaps in Fig. 4 and the dataset-scale argument, i.e., selected using benchmark performance.
axioms (5)
  • domain assumption Cluster assumption: samples located closely in feature space are likely to share the same label.
    Invoked for pseudo-label soft voting (Sec. III-C) and instance contrastive learning on hard samples (Sec. III-E2).
  • domain assumption Deep models learn clean samples faster than noisy ones.
    Borrowed from refs [16], [17] and used to justify progressive easy-to-hard training (Sec. I, Sec. III-B).
  • domain assumption Confidence and prediction entropy of the source pre-trained model identify correctly-classified easy samples in the target domain.
    The entire domain division (Sec. III-B) relies on this correlation; it is not proven and could fail under large domain shift.
  • domain assumption Source and target domains share the same label set of C categories.
    Stated in Sec. III-A and required for class-wise division, pseudo-label assignment, and classification self-training.
  • ad hoc to paper Row-normalized feature dot-product similarities A_ij in Eq. (14) are positive and can be used as probabilities in the log cross-entropy of Eq. (15).
    No softmax or shifting is specified; negative dot products would make log A_ij undefined. This is an unstated implementation assumption.

pith-pipeline@v1.3.0-daily-deepseek · 3431 in / 3676 out tokens · 158515 ms · 2026-08-03T11:41:44.936552+00:00 · methodology

0 comments
read the original abstract

With growing privacy and portability concerns, source-free domain adaptation requires only a source pre-trained model and an unlabeled target domain, allowing for effective adaptation to the target data. Most existing self-training methods focus on selecting and exploiting samples with reliable predictions, often neglecting others. Inspired by the finding that deep models learn clean samples faster than noisy ones, we propose a domain-division based progressive learning method named DPL. Specifically, our approach consists of two alternating stages, each beginning with the division of the target domain into easy-to-adapt and hard-to-adapt subdomains based on adaptation difficulty, followed by neighborhood-based pseudo label assignment. In stage one, we enhance classification accuracy through uncertainty-aware self-training and alignment of corresponding classes between subdomains. Stage two then applies tailored learning strategies to each subdomain, starting with consistency learning on the easy-to-adapt samples and progressing to utilizing local structural information for the more challenging ones, thereby mining the intrinsic properties of the target data. Extensive experiments on several widely used benchmarks validate the effectiveness of our approach, demonstrating superior performance compared to state-of-the-art methods. Our code is available at https://github.com/iamjingli/DPL.

Figures

Figures reproduced from arXiv: 2607.29202 by Jing Li, Meng Zhao, Pan Liu, Qinghua Hu, Shengyong Chen, Wanli Xue.

Figure 1
Figure 1. Figure 1: The overview of our proposed DPL. Its training process comprises of two alternating stages. Each stage begins with [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: t-SNE [70] visualizations of the target domain features for A [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Hyperparameter sensitivity analysis: (a) The effect of [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Hyperparameter sensitivity analysis: The impact of different combinations of [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Prediction accuracy for four Office-Home adaptation tasks. E-acc refers to the accuracy on selected easy samples, while [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Visualization of image regions vital for classification [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

71 extracted references · 2 linked inside Pith

  1. [1]

    Deep learning for ai,

    Y . Bengio, Y . Lecun, and G. Hinton, “Deep learning for ai,” Communications of the ACM, vol. 64, no. 7, pp. 58–65, 2021. 1

    2021

  2. [2]

    Transfer adaptation learning: A decade survey,

    L. Zhang and X. Gao, “Transfer adaptation learning: A decade survey,” IEEE Transactions on Neural Networks and Learning Systems, vol. 35, no. 1, pp. 23–44, 2024. 1, 2

    2024

  3. [3]

    A comprehensive survey on source-free domain adaptation,

    J. Li, Z. Yu, Z. Du, L. Zhu, and H. T. Shen, “A comprehensive survey on source-free domain adaptation,” IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1

    2024

  4. [4]

    Vdm-da: Virtual domain modeling for source data-free domain adaptation,

    J. Tian, J. Zhang, W. Li, and D. Xu, “Vdm-da: Virtual domain modeling for source data-free domain adaptation,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 32, no. 6, pp. 3749–3760, 2021. 1

    2021

  5. [5]

    Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation,

    J. Liang, D. Hu, and J. Feng, “Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation,” in International conference on machine learning. PMLR, 2020, pp. 6028–6039. 1, 2, 4, 6, 7

    2020

  6. [6]

    Denoised maximum classifier discrepancy for source-free unsupervised domain adaptation,

    T. Chu, Y . Liu, J. Deng, W. Li, and L. Duan, “Denoised maximum classifier discrepancy for source-free unsupervised domain adaptation,” in Proceedings of the AAAI conference on artificial intelligence, vol. 36, no. 1, 2022, pp. 472–480. 1

    2022

  7. [7]

    Cleaning noisy labels by nega- tive ensemble learning for source-free unsupervised domain adaptation,

    W. Ahmed, P. Morerio, and V . Murino, “Cleaning noisy labels by nega- tive ensemble learning for source-free unsupervised domain adaptation,” in Proceedings of the IEEE/CVF winter conference on applications of computer vision, 2022, pp. 1616–1625. 1, 2

    2022

  8. [8]

    Self-supervised noisy label learning for source-free unsupervised domain adaptation,

    W. Chen, L. Lin, S. Yang, D. Xie, S. Pu, and Y . Zhuang, “Self-supervised noisy label learning for source-free unsupervised domain adaptation,” in 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2022, pp. 10 185–10 192. 1, 2

    2022

  9. [9]

    Neighborhood-aware mutual information maximization for source-free domain adaptation,

    L. Zhang, Y . Wang, R. Song, M. Zhang, X. Li, and W. Zhang, “Neighborhood-aware mutual information maximization for source-free domain adaptation,” IEEE Transactions on Multimedia, vol. 26, pp. 9564–9574, 2024. 1

    2024

  10. [10]

    Exploiting the intrinsic neighborhood structure for source-free domain adaptation,

    S. Yang, J. Van de Weijer, L. Herranz, and S. Jui, “Exploiting the intrinsic neighborhood structure for source-free domain adaptation,” Advances in neural information processing systems, vol. 34, pp. 29 393– 29 405, 2021. 1, 6, 7

    2021

  11. [11]

    General- ized source-free domain adaptation,

    S. Yang, Y . Wang, J. Van De Weijer, L. Herranz, and S. Jui, “General- ized source-free domain adaptation,” in Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 8978–8987. 1

    2021

  12. [12]

    A survey on negative transfer,

    W. Zhang, L. Deng, L. Zhang, and D. Wu, “A survey on negative transfer,” IEEE/CAA Journal of Automatica Sinica, vol. 10, no. 2, pp. 305–329, 2023. 1

    2023

  13. [13]

    Source data-absent unsupervised domain adaptation through hypothesis transfer and la- beling transfer,

    J. Liang, D. Hu, Y . Wang, R. He, and J. Feng, “Source data-absent unsupervised domain adaptation through hypothesis transfer and la- beling transfer,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 11, pp. 8602–8617, 2021. 1, 2, 6, 7

    2021

  14. [14]

    Mixmatch: A holistic approach to semi-supervised learning,

    D. Berthelot, N. Carlini, I. Goodfellow, N. Papernot, A. Oliver, and C. A. Raffel, “Mixmatch: A holistic approach to semi-supervised learning,” Advances in neural information processing systems, vol. 32, 2019. 1

    2019

  15. [15]

    Intrinsic consistency preservation with adaptively reliable samples for source- free domain adaptation,

    J. Tian, A. E. Saddik, X. Xu, D. Li, Z. Cao, and H. T. Shen, “Intrinsic consistency preservation with adaptively reliable samples for source- free domain adaptation,” IEEE Transactions on Neural Networks and Learning Systems, pp. 1–12, 2024. 1, 2

    2024

  16. [16]

    A closer look at memorization in deep networks,

    D. Arpit, S. Jastrz˛ ebski, N. Ballas, D. Krueger, E. Bengio, M. S. Kanwal, T. Maharaj, A. Fischer, A. Courville, Y . Bengio et al., “A closer look at memorization in deep networks,” in International conference on machine learning. PMLR, 2017, pp. 233–242. 1, 3

    2017

  17. [17]

    Understand- ing deep learning requires rethinking generalization,

    C. Zhang, S. Bengio, M. Hardt, B. Recht, and O. Vinyals, “Understand- ing deep learning requires rethinking generalization,” in International Conference on Learning Representations, 2017. 1, 2, 3

    2017

  18. [18]

    Semi-supervised clustering under a compact-cluster assumption,

    Z. Jiang, Y . Zhan, Q. Mao, and Y . Du, “Semi-supervised clustering under a compact-cluster assumption,” IEEE Transactions on Knowledge and Data Engineering, vol. 35, no. 5, pp. 5244–5256, 2022. 2

    2022

  19. [19]

    A kernel method for the two-sample-problem,

    A. Gretton, K. Borgwardt, M. Rasch, B. Schölkopf, and A. Smola, “A kernel method for the two-sample-problem,” Advances in neural information processing systems, vol. 19, 2006. 2

    2006

  20. [20]

    Wdan: A weighted discriminative adversarial network with dual classifiers for fine-grained open-set do- main adaptation,

    J. Li, L. Yang, Q. Wang, and Q. Hu, “Wdan: A weighted discriminative adversarial network with dual classifiers for fine-grained open-set do- main adaptation,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 33, no. 9, pp. 5133–5147, 2023. 2

    2023

  21. [21]

    Domain prompt tuning via meta relabeling for unsupervised adversarial adaptation,

    X. Jin, C. Lan, W. Zeng, and Z. Chen, “Domain prompt tuning via meta relabeling for unsupervised adversarial adaptation,” IEEE Transactions on Multimedia, vol. 26, pp. 8333–8347, 2024. 2

    2024

  22. [22]

    Ktransgan: Vari- ational inference-based knowledge transfer for unsupervised conditional generative learning,

    M. Azzam, W. Wu, W. Cao, S. Wu, and H.-S. Wong, “Ktransgan: Vari- ational inference-based knowledge transfer for unsupervised conditional generative learning,” IEEE Transactions on Multimedia, vol. 23, pp. 3318–3331, 2020. 2

    2020

  23. [23]

    Cross- domain contrastive learning for unsupervised domain adaptation,

    R. Wang, Z. Wu, Z. Weng, J. Chen, G.-J. Qi, and Y .-G. Jiang, “Cross- domain contrastive learning for unsupervised domain adaptation,” IEEE Transactions on Multimedia, vol. 25, pp. 1665–1673, 2023. 2

    2023

  24. [24]

    Unsupervised domain adaptation via structurally regularized deep clustering,

    H. Tang, K. Chen, and K. Jia, “Unsupervised domain adaptation via structurally regularized deep clustering,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 8725–8735. 2

    2020

  25. [25]

    Universal source-free domain adaptation,

    J. N. Kundu, N. Venkat, and R. V . Babu, “Universal source-free domain adaptation,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 4544–4553. 2

    2020

  26. [26]

    Towards inheritable models for open-set domain adaptation,

    J. N. Kundu, N. Venkat, A. Revanur, and R. V . Babu, “Towards inheritable models for open-set domain adaptation,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 12 376–12 385. 2

    2020

  27. [27]

    Bmd: A general class-balanced multicentric dynamic prototype strategy for source-free domain adaptation,

    S. Qu, G. Chen, J. Zhang, Z. Li, W. He, and D. Tao, “Bmd: A general class-balanced multicentric dynamic prototype strategy for source-free domain adaptation,” in Computer Vision – ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXXIV. Springer, 2022, pp. 165–182. 2

    2022

  28. [28]

    Source- free domain adaptation via target prediction distribution searching,

    S. Tang, A. Chang, F. Zhang, X. Zhu, M. Ye, and C. Zhang, “Source- free domain adaptation via target prediction distribution searching,” International journal of computer vision, vol. 132, no. 3, pp. 654–672,

  29. [29]

    Self-mining the confident prototypes for source-free unsupervised domain adaptation in image segmentation,

    Y . Tian, J. Li, H. Fu, L. Zhu, L. Yu, and L. Wan, “Self-mining the confident prototypes for source-free unsupervised domain adaptation in image segmentation,” IEEE Transactions on Multimedia, vol. 26, pp. 7709–7720, 2024. 2

    2024

  30. [30]

    Divide and contrast: Source-free domain adaptation via adaptive contrastive learning,

    Z. Zhang, W. Chen, H. Cheng, Z. Li, S. Li, L. Lin, and G. Li, “Divide and contrast: Source-free domain adaptation via adaptive contrastive learning,” Advances in neural information processing systems, vol. 35, pp. 5137–5149, 2022. 2

    2022

  31. [31]

    Unsupervised intra- domain adaptation for semantic segmentation through self-supervision,

    F. Pan, I. Shin, F. Rameau, S. Lee, and I. S. Kweon, “Unsupervised intra- domain adaptation for semantic segmentation through self-supervision,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 3764–3773. 2

    2020

  32. [32]

    Adversarial discrim- inative domain adaptation,

    E. Tzeng, J. Hoffman, K. Saenko, and T. Darrell, “Adversarial discrim- inative domain adaptation,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 7167–7176. 2

    2017

  33. [33]

    Mlnet: Mutual learning network with neighborhood invariance for universal domain adaptation,

    Y . Lu, M. Shen, A. J. Ma, X. Xie, and J.-H. Lai, “Mlnet: Mutual learning network with neighborhood invariance for universal domain adaptation,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 4, 2024, pp. 3900–3908. 2

    2024

  34. [34]

    Learning with alignments: Tackling the inter-and intra-domain shifts for cross-multidomain facial expression recognition,

    Y . Yang, L. Wen, X. Zeng, Y . Xu, X. Wu, J. Zhou, and Y . Wang, “Learning with alignments: Tackling the inter-and intra-domain shifts for cross-multidomain facial expression recognition,” in Proceedings of the 32nd ACM International Conference on Multimedia, 2024, pp. 4236–

    2024

  35. [35]

    Divide to adapt: Mitigating confirmation bias for domain adapta- tion of black-box predictors,

    J. Yang, X. Peng, K. Wang, Z. Zhu, J. Feng, L. Xie, and Y . You, “Divide to adapt: Mitigating confirmation bias for domain adapta- tion of black-box predictors,” in International Conference on Learning Representations, 2023. 2

    2023

  36. [36]

    A separation and alignment framework for black-box domain adaptation,

    M. Xia, J. Zhao, G. Lyu, Z. Huang, T. Hu, G. Chen, and H. Wang, “A separation and alignment framework for black-box domain adaptation,” 12 in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 14, 2024, pp. 16 005–16 013. 2, 3, 5

    2024

  37. [37]

    Dividemix: Learning with noisy labels as semi-supervised learning,

    J. Li, R. Socher, and S. C. Hoi, “Dividemix: Learning with noisy labels as semi-supervised learning,” in International Conference on Learning Representations, 2020. 2

    2020

  38. [38]

    Curriculum domain adaptation for semantic segmentation of urban scenes,

    Y . Zhang, P. David, and B. Gong, “Curriculum domain adaptation for semantic segmentation of urban scenes,” in Proceedings of the IEEE international conference on computer vision, 2017, pp. 2020–2030. 2, 3

    2017

  39. [39]

    Model adaptation with synthetic and real data for semantic dense foggy scene understand- ing,

    C. Sakaridis, D. Dai, S. Hecker, and L. Van Gool, “Model adaptation with synthetic and real data for semantic dense foggy scene understand- ing,” in Proceedings of the european conference on computer vision (ECCV), 2018, pp. 687–704. 2, 3

    2018

  40. [40]

    Representation learning with contrastive predictive coding,

    A. v. d. Oord, Y . Li, and O. Vinyals, “Representation learning with contrastive predictive coding,” arXiv preprint arXiv:1807.03748, 2018. 3

    Pith/arXiv arXiv 2018

  41. [41]

    Contrastively smoothed class alignment for unsupervised domain adaptation,

    S. Dai, Y . Cheng, Y . Zhang, Z. Gan, J. Liu, and L. Carin, “Contrastively smoothed class alignment for unsupervised domain adaptation,” in Proceedings of the Asian conference on computer vision, 2020. 3

    2020

  42. [42]

    Model adaptation: Historical contrastive learning for unsupervised domain adaptation without source data,

    J. Huang, D. Guan, A. Xiao, and S. Lu, “Model adaptation: Historical contrastive learning for unsupervised domain adaptation without source data,” Advances in neural information processing systems, vol. 34, pp. 3635–3649, 2021. 3

    2021

  43. [43]

    Class relationship embedded learning for source-free unsupervised domain adaptation,

    Y . Zhang, Z. Wang, and W. He, “Class relationship embedded learning for source-free unsupervised domain adaptation,” in Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, 2023, pp. 7619–7629. 3, 4

    2023

  44. [44]

    A “soft” k-nearest neighbor voting scheme,

    H. B. Mitchell and P. A. Schaefer, “A “soft” k-nearest neighbor voting scheme,” International journal of intelligent systems, vol. 16, no. 4, pp. 459–468, 2001. 4

    2001

  45. [45]

    Towards discriminability and diversity: Batch nuclear-norm maximization un- der label insufficient situations,

    S. Cui, S. Wang, J. Zhuo, L. Li, Q. Huang, and Q. Tian, “Towards discriminability and diversity: Batch nuclear-norm maximization un- der label insufficient situations,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 3941–

    2020

  46. [46]

    Supervised contrastive learn- ing,

    P. Khosla, P. Teterwak, C. Wang, A. Sarna, Y . Tian, P. Isola, A. Maschinot, C. Liu, and D. Krishnan, “Supervised contrastive learn- ing,” Advances in neural information processing systems, vol. 33, pp. 18 661–18 673, 2020. 5

    2020

  47. [47]

    Enhancing multi-source open-set domain adaptation through nearest neighbor classification with self-supervised vision transformer,

    J. Li, L. Yang, and Q. Hu, “Enhancing multi-source open-set domain adaptation through nearest neighbor classification with self-supervised vision transformer,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 34, no. 4, pp. 2648–2662, 2024. 5

    2024

  48. [48]

    Probabilistic contrastive learning for domain adaptation,

    J. Li, Y . Zhang, Z. Wang, S. Hou, K. Tu, and M. Zhang, “Probabilistic contrastive learning for domain adaptation,” in Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, IJCAI-24, 8 2024, pp. 1001–1009. 5

    2024

  49. [49]

    Adapting visual cate- gory models to new domains,

    K. Saenko, B. Kulis, M. Fritz, and T. Darrell, “Adapting visual cate- gory models to new domains,” in Computer Vision–ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part IV 11. Springer, 2010, pp. 213–226. 6

    2010

  50. [50]

    Deep hashing network for unsupervised domain adaptation,

    H. Venkateswara, J. Eusebio, S. Chakraborty, and S. Panchanathan, “Deep hashing network for unsupervised domain adaptation,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 5018–5027. 6

    2017

  51. [51]

    Visda: The visual domain adaptation challenge,

    X. Peng, B. Usman, N. Kaushik, J. Hoffman, D. Wang, and K. Saenko, “Visda: The visual domain adaptation challenge,” arXiv preprint arXiv:1710.06924, 2017. 6

    Pith/arXiv arXiv 2017

  52. [52]

    Conditional adversarial domain adaptation,

    M. Long, Z. Cao, J. Wang, and M. I. Jordan, “Conditional adversarial domain adaptation,” Advances in neural information processing systems, vol. 31, 2018. 6, 7

    2018

  53. [53]

    Larger norm more transferable: An adaptive feature norm approach for unsupervised domain adaptation,

    R. Xu, G. Li, J. Yang, and L. Lin, “Larger norm more transferable: An adaptive feature norm approach for unsupervised domain adaptation,” in Proceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 1426–1435. 6, 7

    2019

  54. [54]

    Implicit class-conditioned domain alignment for unsupervised domain adaptation,

    X. Jiang, Q. Lao, S. Matwin, and M. Havaei, “Implicit class-conditioned domain alignment for unsupervised domain adaptation,” in International conference on machine learning. PMLR, 2020, pp. 4816–4827. 6, 7

    2020

  55. [55]

    Maximum structural generation discrep- ancy for unsupervised domain adaptation,

    H. Xia, T. Jing, and Z. Ding, “Maximum structural generation discrep- ancy for unsupervised domain adaptation,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 3, pp. 3434–3445, 2022. 6, 7

    2022

  56. [56]

    Adaptive mutual learning for unsupervised domain adaptation,

    L. Zhou, S. Xiao, M. Ye, X. Zhu, and S. Li, “Adaptive mutual learning for unsupervised domain adaptation,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 33, no. 11, pp. 6622–6634, 2023. 6, 7

    2023

  57. [57]

    Multiview latent space learning with progressively fine-tuned deep features for unsupervised domain adaptation,

    C. Zhu, Q. Wang, Y . Xie, and S. Xu, “Multiview latent space learning with progressively fine-tuned deep features for unsupervised domain adaptation,” Information Sciences, vol. 662, p. 120223, 2024. 6, 7

    2024

  58. [58]

    Exploring domain-invariant parameters for source free domain adaptation,

    F. Wang, Z. Han, Y . Gong, and Y . Yin, “Exploring domain-invariant parameters for source free domain adaptation,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 7151–7160. 6, 7

    2022

  59. [59]

    Attracting and dispersing: A simple approach for source-free domain adaptation,

    S. Yang, S. Jui, and J. van de Weijer, “Attracting and dispersing: A simple approach for source-free domain adaptation,” Advances in neural information processing systems, vol. 35, pp. 5802–5815, 2022. 6, 7

    2022

  60. [60]

    Conmix for source- free single and multi-target domain adaptation,

    V . Kumar, R. Lal, H. Patil, and A. Chakraborty, “Conmix for source- free single and multi-target domain adaptation,” in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2023, pp. 4178–4188. 6, 7

    2023

  61. [61]

    Uncertainty- induced transferability representation for source-free unsupervised do- main adaptation,

    J. Pei, Z. Jiang, A. Men, L. Chen, Y . Liu, and Q. Chen, “Uncertainty- induced transferability representation for source-free unsupervised do- main adaptation,” IEEE Transactions on Image Processing, vol. 32, pp. 2033–2048, 2023. 6, 7

    2033

  62. [62]

    Unified multi-level neighbor clustering for source-free unsupervised domain adaptation,

    Y . Xiao, G. Xiao, and H. Chen, “Unified multi-level neighbor clustering for source-free unsupervised domain adaptation,” Pattern Recognition, vol. 153, p. 110533, 2024. 6, 7

    2024

  63. [63]

    Model adap- tation through hypothesis transfer with gradual knowledge distillation,

    S. Tang, Y . Shi, Z. Ma, J. Li, J. Lyu, Q. Li, and J. Zhang, “Model adap- tation through hypothesis transfer with gradual knowledge distillation,” in 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2021, pp. 5679–5685. 7

    2021

  64. [64]

    Divergence- agnostic unsupervised domain adaptation by adversarial attacks,

    J. Li, Z. Du, L. Zhu, Z. Ding, K. Lu, and H. T. Shen, “Divergence- agnostic unsupervised domain adaptation by adversarial attacks,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 11, pp. 8196–8211, 2021. 7

    2021

  65. [65]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778. 6

    2016

  66. [66]

    Weight normalization: A simple repa- rameterization to accelerate training of deep neural networks,

    T. Salimans and D. P. Kingma, “Weight normalization: A simple repa- rameterization to accelerate training of deep neural networks,” Advances in neural information processing systems, vol. 29, 2016. 6

    2016

  67. [67]

    Imagenet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: A large-scale hierarchical image database,” in 2009 IEEE conference on computer vision and pattern recognition. Ieee, 2009, pp. 248–255. 7

    2009

  68. [68]

    Minimum class confusion for versatile domain adaptation,

    Y . Jin, X. Wang, M. Long, and J. Wang, “Minimum class confusion for versatile domain adaptation,” in Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXI 16. Springer, 2020, pp. 464–480. 7

    2020

  69. [69]

    Spectral unsupervised domain adaptation for visual recognition,

    J. Zhang, J. Huang, Z. Tian, and S. Lu, “Spectral unsupervised domain adaptation for visual recognition,” in Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, 2022, pp. 9829–

    2022

  70. [70]

    Visualizing data using t-sne

    L. Van der Maaten and G. Hinton, “Visualizing data using t-sne.” Journal of machine learning research, vol. 9, no. 11, 2008. 8, 9

    2008

  71. [71]

    Grad-cam: Visual explanations from deep networks via gradient-based localization,

    R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Batra, “Grad-cam: Visual explanations from deep networks via gradient-based localization,” in Proceedings of the IEEE international conference on computer vision, 2017, pp. 618–626. 9, 10

    2017