Pith. sign in

REVIEW 3 major objections 5 minor 37 references

PCaM: A Progressive Focus Attention-Based Information Fusion Method for Improving Vision Transformer Domain Adaptation

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that foreground object mismatch—where the same object occupies different sizes and positions across domains—is a hidden blocker in ViT-based domain adaptation, and that a plug-in cross-attention module which…

desk verdict PCaM has a real empirical package, but its own Eq. (10) makes the central focus loss repulsive, so the stated mechanism cannot explain the reported gains. read the letter →

arxiv 2506.17232 v1 pith:CRPD7MZW submitted 2025-05-27 cs.LG cs.AIcs.CV

Zelin Zang , Fei Wang , Liangyu Li , Jinlin Wu , Chunshui Zhao , Zhen Lei , Baigui Sun This is my paper · ORCID
classification cs.LGcs.AIcs.CV
keywords UnsupervisedDomainAdaptationVisionTransformerForegroundObjectMismatchCross-AttentionAttentionRolloutFeatureRefinementProgressiveFocusLossCross-DomainInformationFusion
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that ViT-based unsupervised domain adaptation fails when the foreground object occupies very different fractions of the image in source and target domains. It introduces PCaM, a plug-in module that uses cross-attention rollout to locate the foreground, crops and rescales it, and applies a progressively focused loss to keep attention on that region. On four benchmarks the module improves a CDTrans baseline, reaching 91.4% average accuracy on VisDA-2017 and 47.2% on DomainNet. The paper's intended contribution is that attention-guided foreground fusion, not stronger feature alignment alone, is what closes the remaining domain gap.

What carries the argument

The central object is the cross-attention rollout $\overline{AR}^{st,l}_{i,j}$, a recursive accumulation of patch-wise source-target attention over transformer layers that estimates which target patches correspond to a source foreground patch. It drives two operations: box identification and interpolation (feature refinement) that crop and rescale the attended region, and the progressively focused loss $L_{PF}$, which weights the squared difference between each attention value and the attention center of mass by inverse distance, intending to pull attention toward the foreground center. The loss is added to CDTrans's classification and distillation losses, and a center-aware pseudo-labeling filter decides which source-target pairs are kept.

What would settle it

Compute the gradient of Eq. (10) on a toy attention map with a single hot spot: with the printed negative sign, gradient descent will increase already-high values and decrease already-low ones, so the alleged pulling toward the center cannot occur unless the sign is reversed or the loss is redefined. Separately, compare attention-rollout bounding boxes to ground-truth object masks in the target domain; if the boxes do not cover the same object as in the source, the foreground-alignment premise fails.

Watch

Extended reading notes

Core claim

PCaM is the claim that foreground object mismatch—size and spatial-distribution differences of the same object across domains—is a key hidden obstacle in ViT-based UDA, and that filtering background during cross-attention fixes it. The mechanism progressively refines attention maps through three steps: attention rollout accumulates patch-level cross-domain relevance, feature refinement crops and resizes the box the rollout highlights, and a progressively focused loss concentrates attention on a continuous local region. The paper reports state-of-the-art results on VisDA-2017 (91.4%), DomainNet (47.2%), Office-Home (81.3%), and AID→NWPU (88.6%), with the largest class-level gains on exactly the categories where foreground ratios differ most between domains.

Load-bearing premise

The mechanism presumes that the regions highlighted by attention rollout are the same semantic foreground in both domains, and that minimizing Eq. (10) as printed moves attention toward the center of mass rather than away from it.

Editorial extensions

If this is right

  • PCaM's 91.4% average on VisDA-2017, if reproduced, makes attention-guided foreground fusion the current best ViT-based UDA result on that benchmark.
  • On classes with small foreground objects in cluttered scenes—sktbrd, bcycl, and truck—the reported gains of 8.6, 6.8, and 7.5 points over CDTrans show that the method targets the exact failure mode it identifies.
  • On DomainNet, the 47.2% average with +2.0 over CDTrans suggests the mechanism scales to large, many-class benchmarks rather than only small ones.
  • At inference PCaM is discarded; test images are fed directly to the ViT, so the improved accuracy comes at no inference-time cost.
  • The remote sensing improvement (88.6% on AID→NWPU, +4.9 over the CDTrans baseline) indicates the method transfers to aerial imagery where foreground/background structure differs substantially.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A sign-corrected version of $L_{PF}$ would make the focusing mechanism directly testable as a general attention regularizer for any transformer that must attend to a compact foreground, such as video object segmentation or visual grounding.
  • The per-class gains on sktbrd, bcycl, and truck suggest that a simple foreground-ratio mismatch statistic computed before training could predict which classes will benefit most from PCaM.
  • If the gains survive a synthetic test where foreground objects are identical across domains and only backgrounds differ, the FOM story would be confirmed; if not, the improvement may come from cropping or regularization rather than cross-domain fusion.
  • The dual filtering of pseudo-label agreement and feature similarity could be reused in other UDA pipelines independently of the foreground-focus loss.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes PCaM, a plug-in module for ViT-based unsupervised domain adaptation (UDA) that addresses a hypothesized 'Foreground Object Mismatch' (FOM) issue. PCaM uses attention rollout to localize foreground regions, a feature-refinement step that crops and resizes those regions, and a 'progressively focused' loss (LPF) intended to concentrate attention toward the center of mass. The method is evaluated on Office-Home, VisDA-2017, DomainNet, and AID-to-NWPU, reporting improvements over CDTrans and several other baselines, with ablation and robustness experiments.

Significance. If the described mechanism were correct, PCaM would be an appealingly simple and architecture-agnostic ingredient for ViT UDA, with broad empirical gains and a plausible story for why foreground/background mismatch hurts cross-domain attention. The paper also attempts theoretical justifications in Appendices B-D, which is a positive sign. However, the central loss in Eq. (10) is written with a sign that makes it repulsive rather than focusing, and Appendix C's own gradient analysis confirms the repulsive direction while mislabeling it as 'pulling'. Appendix B's 'foreground aggregation' result assumes the conclusion it claims to prove. These are load-bearing flaws: they concern the very mechanism that distinguishes PCaM from its base method, CDTrans. The empirical results are extensive, but the attribution of the reported gains to the proposed progressive-focusing mechanism is not supported by the manuscript as written.

major comments (3)
  1. [Eq. (10) and Appendix C] The loss in Eq. (10) is the negative of a weighted squared distance to the center: LPF = -Σ ||AR_{J(m,n)} - AR_{J(⌊m_c⌋,⌊n_c⌋)}||^2 / sqrt((m-m_c)^2 + (n-n_c)^2). Minimizing this loss maximizes the weighted squared deviation of each attention value from the center value, which is the opposite of the focusing behavior claimed in Section 3.4. Appendix C derives ∂LPF/∂A_{m,n} = -2(A_{m,n} - A_c)/d_{m,n} and calls this a 'pulling' effect, but under gradient descent this gradient is repulsive: values below A_c are pushed lower and values above A_c are pushed higher, with the strongest effect near the center. The appendix acknowledges omitted chain-rule terms through A_c and (m_c, n_c) but does not show they reverse the dominant sign. Since LPF is the component that differentiates PCaM from CDTrans, the reported gains cannot be attributed to the described mechanism unless the implemented code used the opposite sign, which is not stated anywhere in the manuscript.
  2. [Appendix B, Proposition 1] Proposition 1 ('Foreground Aggregation Trend') is circular. It assumes that the source query patch is in the foreground and that the source-to-target attention maps concentrate on target foreground patches, and then concludes that the normalized attention rollout assigns higher values to foreground patches. This is a restatement of the premise, not a proof. Section 3.3 also states that attention rollout identifies regions 'postulating these regions to correspond to foreground objects'; the Appendix B claim that the rollout 'formally prove[s]' foreground aggregation is therefore not supported. The theoretical support for the central localization mechanism is an assumption, not a result.
  3. [Table 4] The ablation study does not isolate the contribution of the PF loss. Table 4 reports Baseline (88.4), +Crop Fore. Patches (88.5), +Weight Patches (88.5), +Det (88.9), +Det+AR (89.7), +Det+AR+FR (90.2), and PCaM Full (91.4), but the text never defines which components are included in 'Det', 'AR', and 'FR', and the LPF term of Eq. (10) is never separately ablated. The 1.2-point improvement from '+Det+AR+FR' to 'PCaM (Full)' is the only evidence that could be attributed to the loss, but without a controlled ablation that toggles LPF alone, the claim that the progressively focused loss is responsible for the gains is not demonstrated.
minor comments (5)
  1. [§4.3 and Table 2] The reported margin over the previous best method is numerically inconsistent: Table 2 lists C-SFTrans at 89.6 and PCaM at 91.4, a difference of 1.8 points, while the text in §4.3 says 'outperforming the previous best method C-SFTrans (2024) by 1.1%' and the Table 2 caption says 'outperforms the second best method by 1.1%'. If the intended second best is DOT at 90.3, the caption should identify it explicitly.
  2. [§4.5] Section 4.5 references 'Fig.X and TableX' for the pseudo-label robustness results; these placeholders are never resolved and should be replaced with actual figure and table numbers.
  3. [§4.2/§4.3] The computational cost paragraph says the results are 'illustrated in Fig. 5', but Fig. 5 shows attention rollout visualizations; the cost comparison appears in Table 5, so the cross-reference is wrong.
  4. [Appendix E] Appendix E reports PCaM at 90.8% on VisDA-2017 while Table 2 reports 91.4%; the two numbers should be reconciled or the difference explained (e.g., different training settings).
  5. [§2 and §4.4] The text contains informal or typographical errors: 'more related words are in the appendix' in §2, 'Discustions' in the §4.4 heading, and 'Comparision' in §4.3. These should be corrected.

Circularity Check

2 steps flagged · score 6.0 of 10

Benchmark gains are externally evaluated and not circular, but the paper's theoretical core is: Appendix B's foreground-aggregation theorem assumes its own conclusion, and Eq. (10)'s PF loss as printed makes the opposite of the claimed focusing effect by construction.

  1. other [Appendix B, Proposition 1 (Foreground Aggregation Trend); promised in Section 3.3, Attention Rollout]
    "Proposition 1 (Foreground Aggregation Trend). If the source query patch i corresponds to the foreground region, and the source-to-target attention maps σ(z^{s,l}_i · z^{t,l}_j) concentrate on foreground patches j in target, then ˆAR^l_{i,j} tends to assign higher values to foreground j than background j."

    The proposition's conclusion, that attention rollout assigns higher values to foreground patches, is exactly its premise, that the source-to-target attention maps concentrate on foreground patches. The proof sketch only restates the premise: a foreground query aligns with target foreground patches, so softmax yields higher weights there, and AR aggregates this emphasis. The boundedness and Cesàro-convergence lemmas do not establish any foreground selectivity. Thus the main text's claim that AR 'formally prove[s] ... consistent foreground aggregation behavior' is an assumption restated as a theorem, not a derivation.

  2. self definitional [Section 3.4, Eq. (10), and Appendix C, Proposition 2]
    "L_PF = −∑_l ∑_{m,n=1}^N ‖AR^{st,l}_{J(m,n)} − AR^{st,l}_{J(⌊mlc⌋,⌊nlc⌋)}‖_2^2 / √((m−m_lc)^2+(n−n_lc)^2) ... ∂L_PF/∂A_{m,n} = −2·(A_{m,n} − A_c)/d_{m,n} + additional terms ... The first term dominates and exhibits a 'pulling' effect: when A_{m,n} > A_c, the gradient is negative, reducing A_{m,n}; when A_{m,n} < A_c, the gradient is positive, increasing A_{m,n}."

    Because of the leading minus sign, minimizing the printed L_PF is by construction equivalent to maximizing the weighted squared deviation of attention values from the center value. The gradient in Appendix C is the gradient of that negative quadratic; gradient descent follows its negative, so A_{m,n} moves in the direction +2(A_{m,n}−A_c)/d_{m,n}, i.e., away from the center, and the inverse-distance weighting makes this repulsion strongest near the center. The claimed 'pulling' effect is therefore not a consequence of Eq. (10) as written; the focusing behavior requires an unstated sign flip or a different loss. The paper's central mechanism is thus an artifact of the printed definition rather than a derived property.

full rationale

The reported accuracy gains on VisDA-2017, Office-Home, DomainNet, and AID→NWPU are comparisons against external baselines with shared DeiT-B backbones, so the empirical headline is not circular and is independently checkable. No load-bearing self-citation chain or imported uniqueness theorem appears. However, the theoretical scaffolding for the method's two novel components is not self-contained. Appendix B's Proposition 1 is a circular proof: its conclusion (attention rollout emphasizes foreground) is assumed in its premise (attention maps concentrate on foreground patches). The boundedness and convergence lemmas prove only numerical stability, not semantic foreground aggregation. Separately, the progressively focused loss in Eq. (10) is defined with a negative sign, so minimizing it maximizes, not minimizes, the weighted squared deviation from the attention center of mass; Appendix C's gradient computation confirms this but misreads the descent direction, calling repulsion a 'pulling' effect. This is a self-definitional inconsistency: the claimed focusing mechanism is the opposite of what the printed loss does by construction. Together these issues make the paper's derivation chain partially circular, while leaving the external benchmark results as independent empirical evidence. Score 6 reflects one prediction (foreground aggregation) reducing to its premise and one loss-defined behavior reducing to its opposite.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The method depends on several domain assumptions about attention semantics and pseudo-label quality. No new entities are introduced. The only fitted free parameters are the box threshold and the PF loss weight, both selected on the evaluation benchmarks. The theoretical appendices do not supply independent grounding.

free parameters (2)
  • Box threshold beta = 0.05
    Controls the cropping box size in Eq. (6). Chosen from a parameter sweep on VisDA-2017 (Fig. 7); the paper notes performance degrades when beta exceeds 0.2.
  • PF loss weight = 1.0 in main text; 0.2-0.5 for VisDA and 1.0 for DomainNet per Appendix J
    Appendix G says all loss weights are fixed at 1.0, but Appendix J reports tuning the LPF weight per dataset and says different values give the best results. This is a free parameter selected using the evaluation benchmarks.
assumptions (3)
  • domain assumption High attention rollout values mark the same foreground object in source and target images.
    Introduced in Section 3.3 as 'postulating these regions to correspond to foreground objects' and used to define cropping boxes. Appendix B Proposition 1 assumes rather than proves this.
  • domain assumption Feature-similar paired samples with agreeing pseudo-labels are same-class pairs that improve adaptation.
    Data pairing in Section 3.2 relies on feature-space distance and pseudo-label consistency. If pairs are wrong, the cross-attention boxes and refinement inherit the error.
  • domain assumption Cropping and resizing the attention box preserves discriminative foreground semantics.
    Feature Refinement in Section 3.3 interpolates the box to the original size. The paper provides no evidence that resizing does not distort class-discriminative shape or scale information.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PCaM: A Progressive Focus Attention-Based Information Fusion Method for Improving Vision Transformer Domain Adaptation." pith.science (2026). https://pith.science/paper/CRPD7MZW

@misc{pith2026250617232,
  author       = {Pith},
  title        = {Pith review of: PCaM: A Progressive Focus Attention-Based Information Fusion Method for Improving Vision Transformer Domain Adaptation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CRPD7MZW}},
  note         = {Machine review of arXiv:2506.17232}
}
read the original abstract

Unsupervised Domain Adaptation (UDA) aims to transfer knowledge from a labeled source domain to an unlabeled target domain. Recent UDA methods based on Vision Transformers (ViTs) have achieved strong performance through attention-based feature alignment. However, we identify a key limitation: foreground object mismatch, where the discrepancy in foreground object size and spatial distribution across domains weakens attention consistency and hampers effective domain alignment. To address this issue, we propose the Progressive Focus Cross-Attention Mechanism (PCaM), which progressively filters out background information during cross-attention, allowing the model to focus on and fuse discriminative foreground semantics across domains. We further introduce an attentional guidance loss that explicitly directs attention toward task-relevant regions, enhancing cross-domain attention consistency. PCaM is lightweight, architecture-agnostic, and easy to integrate into existing ViT-based UDA pipelines. Extensive experiments on Office-Home, DomainNet, VisDA-2017, and remote sensing datasets demonstrate that PCaM significantly improves adaptation performance and achieves new state-of-the-art results, validating the effectiveness of attention-guided foreground fusion for domain adaptation.

Figures

Figures reproduced from arXiv: 2506.17232 by the authors.

Figure 1
Figure 1. Highlighting the Foreground Object Mismatch Issue and the En￾hancement Achieved with PCaM in the VisDA Dataset. Part (a) presents bar plots indicating foreground ratios (defined by the proportion of pixels encircled in red to the total pixel count) for individual categories across both source and target domains. Part (b) depicts the bar plots representing the accuracy of each category. In contrast, part (c) showcase… view at source ↗
Figure 2
Figure 2. Core concept of PCaM. PCaM addresses the FOM issue by utilizing the cross-attention module to isolate the foreground and filter out unrelated backgrounds. As illustrated, bicycle samples from two distinct domains are harmonized to comparable fore￾ground scales using PCaM, facilitating improved transfer learning. Within the attention map, a darker hue indicates heightened attention. flattened embedding zj ∈ RD throug… view at source ↗
Figure 3
Figure 3. Framework of Proposed PCaM. Our PCaM framework comprises Attention Rollout and PF Loss components. Attention Rollout guides feature cropping, while PF Loss encourages the model to focus on a continuous local region. The source/target feature contain the information of the source/target domain, the fusion feature is the concatenation of the source and target feature. discarded. In addition, to further ensure the qual… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Confusion matrix on DomainNet. Comparison on confusion matrix with DsiT, DOT, and CDTrains on DomainNet. PCaM achieves the best performance on the DomainNet dataset. PCaM outperforms the second best method by 2.0%. the attention map, [a s,l < , a s,l > , as,l ∧ , as,l …
Figure 5
Figure 5. Figure 5: Visualization of attention rollout on the Office-Home. Here, Xs and Xt represent the image data from the source and target domains respectively. ARss,L visualizes the feature map of the last layer of self-attention in the source domain, whereas ARst,L visualizes the fe…
Figure 6
Figure 6. Figure 6: Exploratory experiment of PCaM. (a) Barplot and line plot illustrating performance as the number of training epochs increases. (b) The cross-attention mecha￾nism of PCaM highlights regions corresponding to different epoch frames. categories sensitive to spatial mismatc…
Figure 7
Figure 7. Figure 7: Parameter analysis. The parameter analysis of parameter β (Beta) in Eq. (6) on VisDA-2017. The performance of PCaM is stable when β is around 0.05 [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]
Figure 7
Figure 7. Figure 7: PCaM Beyond directly Cropping Foreground & Weight Patches. As shown in Table. 4, PCaM mitigates the FOM problem by adapting to the foreground by gradually focusing during training, in order to demonstrate the novelty of this scheme. We tested two more direct schemes, C…
Figure 8
Figure 8. Figure 8: Scatter visualization of CDTrans and PCaM: t-SNE visualization of domain alignment effects. PCaM shows better domain alignment than CDTrans [PITH_FULL_IMAGE:figures/full_fig_p019_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 35 canonical work pages

  1. [1]

    Self- supervised representation learning for geospatial objects: A survey.Informa- tion Fusion, page 103265, 2025

    Yile Chen, Weiming Huang, Kaiqi Zhao, Yue Jiang, and Gao Cong. Self- supervised representation learning for geospatial objects: A survey.Informa- tion Fusion, page 103265, 2025

    work page 2025

  2. [2]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy et al. An image is worth 16x16 words: Transformers for image recognition at scale. InICLR, 2020

    work page 2020

  3. [3]

    Cross-domain gradi- ent discrepancy minimization for unsupervised domain adaptation

    Zhekai Du, Jingjing Li, Hongzu Su, Lei Zhu, and Ke Lu. Cross-domain gradi- ent discrepancy minimization for unsupervised domain adaptation. InCVPR, pages 3937–3946, 2021

    work page 2021

  4. [4]

    Learning to detect open classes for universal domain adap- tation

    Bo Fu and et.al. Learning to detect open classes for universal domain adap- tation. InECCV, number 567–583, 2020

    work page 2020

  5. [5]

    Mic: Masked image consistency for context-enhanced domain adaptation

    Lukas Hoyer, Dengxin Dai, Haoran Wang, and Luc Van Gool. Mic: Masked image consistency for context-enhanced domain adaptation. InCVPR, pages 11721–11732, 2023

    work page 2023

  6. [6]

    Multi- source domain adaptation for panoramic semantic segmentation.Information Fusion, 117:102909, 2025

    Jing Jiang, Sicheng Zhao, Jiankun Zhu, Wenbo Tang, Zhaopan Xu, Jidong Yang, Guoping Liu, Tengfei Xing, Pengfei Xu, and Hongxun Yao. Multi- source domain adaptation for panoramic semantic segmentation.Information Fusion, 117:102909, 2025

    work page 2025

  7. [7]

    Patch-mix transformer for unsupervised domain adaptation: A game perspective

    Lin Wang Jinjing Zhu, Haotian Bai. Patch-mix transformer for unsupervised domain adaptation: A game perspective. 2023

    work page 2023

  8. [8]

    Crossfuse: A novel cross attention mechanism based infrared and visible image fusion approach.Information Fusion, 103:102147, 2024

    Hui Li and Xiao-Jun Wu. Crossfuse: A novel cross attention mechanism based infrared and visible image fusion approach.Information Fusion, 103:102147, 2024. 20

    work page 2024

Show all 37 references
  1. [9]

    Cross-domain adaptive clustering for semi-supervised domain adaptation

    Jichang Li, Guanbin Li, Yemin Shi, and Yizhou Yu. Cross-domain adaptive clustering for semi-supervised domain adaptation. InCVPR, 2021

    2021

  2. [10]

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

    Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. InICML, pages 6028–6039. PMLR, 2020

    2020

  3. [11]

    Foregroundguidanceandmulti- layer feature fusion for unsupervised object discovery with transformers

    ZhiweiLin, ZengyuYang, andYongtaoWang. Foregroundguidanceandmulti- layer feature fusion for unsupervised object discovery with transformers. In W ACV, pages 4043–4053, 2023

    2023

  4. [12]

    F2net: Learning to focus on the foreground for unsupervised video object segmentation.AAAI, 35(3):2109–2117, 2021

    Daizong Liu, Dongdong Yu, Changhu Wang, and Pan Zhou. F2net: Learning to focus on the foreground for unsupervised video object segmentation.AAAI, 35(3):2109–2117, 2021

    2021

  5. [13]

    Explicitly fusing plug-and-play guidance of source prototype into target sub- space for domain adaptation.Information Fusion, 123:103197, 2025

    Hao Luo, Zhiqiang Tian, Panpan Jiao, Meiqin Liu, Shaoyi Du, and Kai Nan. Explicitly fusing plug-and-play guidance of source prototype into target sub- space for domain adaptation.Information Fusion, 123:103197, 2025

    2025

  6. [14]

    Explicitly fusing plug-and-play guidance of source prototype into target sub- space for domain adaptation.Information Fusion, page 103197, 2025

    Hao Luo, Zhiqiang Tian, Panpan Jiao, Meiqin Liu, Shaoyi Du, and Kai Nan. Explicitly fusing plug-and-play guidance of source prototype into target sub- space for domain adaptation.Information Fusion, page 103197, 2025

    2025

  7. [15]

    Making the best of both worlds: A domain-oriented transformer for unsupervised domain adaptation

    Wenxuan Ma and Wei Li. Making the best of both worlds: A domain-oriented transformer for unsupervised domain adaptation. InACM MM, pages 5620– 5629, 2022

    2022

  8. [16]

    Fixbi: Bridging domain spaces for unsupervised domain adaptation

    Jaemin Na, Heechul Jung, Hyung Jin Chang, and Wonjun Hwang. Fixbi: Bridging domain spaces for unsupervised domain adaptation. InCVPR, 2021

    2021

  9. [17]

    Moment matching for multi-source domain adaptation.Int

    Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation.Int. Conf. Comput. Vis., 2019

    2019

  10. [18]

    Visda: The visual domain adaptation challenge.arXiv preprint arXiv:1710.06924, 2017

    Xingchao Peng, Ben Usman, Neela Kaushik, Judy Hoffman, Dequan Wang, and Kate Saenko. Visda: The visual domain adaptation challenge.arXiv preprint arXiv:1710.06924, 2017

    2017 arXiv

  11. [19]

    Domain-specificity inducing transformers for source-free domain adaptation

    Sunandini Sanyal. Domain-specificity inducing transformers for source-free domain adaptation. InICCV, June 2023

    2023

  12. [20]

    Aligning non-causal factors for transformer-based source- free domain adaptation

    Sunandini Sanyal. Aligning non-causal factors for transformer-based source- free domain adaptation. InProceedings of the W ACV, June 2024. 21

    2024

  13. [21]

    On the origin of implicit regularization in stochastic gradient descent.arXiv preprint arXiv:2101.12176, 2024

    Samuel L Smith, Benoit Dherin, David GT Barrett, and Soham De. On the origin of implicit regularization in stochastic gradient descent.arXiv preprint arXiv:2101.12176, 2024

    2024 arXiv

  14. [22]

    Training data-efficient image transformers & distillation through attention

    Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. InICML, pages 10347–10357. PMLR, 2021

    2021

  15. [23]

    Zico Kolter, Louis- Philippe Morency, and Ruslan Salakhutdinov

    Yao-Hung Hubert Tsai, Shaojie Bai, Paul Pu Liang, J. Zico Kolter, Louis- Philippe Morency, and Ruslan Salakhutdinov. Multimodal transformer for unaligned multimodal language sequences. InACL, 2019

    2019

  16. [24]

    Deep hashing network for unsupervised domain adaptation

    Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. IEEE Conf. Comput. Vis. Pattern Recog., 2017

    2017

  17. [25]

    Nwpu-crowd: A large-scale benchmark for crowd counting and localization.TPAMI, 43(6):2141–2149, 2020

    Qi Wang, Junyu Gao, Wei Lin, and Xuelong Li. Nwpu-crowd: A large-scale benchmark for crowd counting and localization.TPAMI, 43(6):2141–2149, 2020

    2020

  18. [26]

    Multidimensional fusion of frequency and spatial domain information for enhanced camouflaged object detection.Information Fusion, 117:102871, 2025

    Tingran Wang, Zaiyang Yu, Jianwei Fang, Jinlong Xie, Feng Yang, Huang Zhang, Liping Zhang, Minghua Du, Lusi Li, and Xin Ning. Multidimensional fusion of frequency and spatial domain information for enhanced camouflaged object detection.Information Fusion, 117:102871, 2025

    2025

  19. [27]

    Multi-teacher self-distillation based on adaptive weighting and activation pattern for enhancing lightweight arrhythmia recognition.Information Fusion, 122:103178, 2025

    Zhongyu Wang, Caiyun Ma, Shuo Zhang, Minghui Zhao, Yu Liu, Lina Zhao, Tingting Zhu, Jianqing Li, and Chengyu Liu. Multi-teacher self-distillation based on adaptive weighting and activation pattern for enhancing lightweight arrhythmia recognition.Information Fusion, 122:103178, 2025

    2025

  20. [28]

    Aid: A benchmark data set for perfor- mance evaluation of aerial scene classification.TGRS, 55(7):3965–3981, 2017

    Gui-Song Xia, Jingwen Hu, Fan Hu, Baoguang Shi, Xiang Bai, Yanfei Zhong, Liangpei Zhang, and Xiaoqiang Lu. Aid: A benchmark data set for perfor- mance evaluation of aerial scene classification.TGRS, 55(7):3965–3981, 2017

    2017

  21. [29]

    Universal domain adaptation for remote sensing image scene classification.TGRS, 61, 2023

    Qingsong Xu and et.al. Universal domain adaptation for remote sensing image scene classification.TGRS, 61, 2023

    2023

  22. [30]

    Cdtrans: Cross-domain transformer for unsupervised domain adaptation

    Tongkun Xu, Weihua Chen, WANG Pichao, Fan Wang, Hao Li, and Rong Jin. Cdtrans: Cross-domain transformer for unsupervised domain adaptation. InInternational Conference on Learning Representations, 2022

    2022

  23. [31]

    Ccin-sa: Composite cross modal interaction network with attention enhancement for multimodal sentiment analysis.Information Fusion, page 103230, 2025

    Li Yang, Junhong Zhong, Teng Wen, and Yuan Liao. Ccin-sa: Composite cross modal interaction network with attention enhancement for multimodal sentiment analysis.Information Fusion, page 103230, 2025. 22

    2025

  24. [32]

    Pseudo-margin-based universal domain adaptation

    Yueming Yin and et.al. Pseudo-margin-based universal domain adaptation. InKnowledge-Based Systems, volume 229, June 2021

    2021

  25. [33]

    Universal domain adaptation

    Kaichao You and et.al. Universal domain adaptation. InCVPR, number 2720-2729, 2019

    2019

  26. [34]

    Dlme: Deep local-flatness manifold embedding

    Zelin Zang, Siyuan Li, Di Wu, Ge Wang, Kai Wang, Lei Shang, Baigui Sun, Hao Li, and Stan Z Li. Dlme: Deep local-flatness manifold embedding. In European Conference on Computer Vision, pages 576–592. Springer, 2022

    2022

  27. [35]

    Zelin Zang, Lei Shang, Senqiao Yang, Fei Wang, Baigui Sun, Xuansong Xie, and Stan Z. Li. Boosting novel category discovery over domains with soft contrastive learning and all-in-one classifier, 2023

    2023

  28. [36]

    Free lunch for domain adversarial training: En- vironment label smoothing

    YiFan Zhang, Xue Wang, Jian Liang, Zhang Zhang, Liang Wang, and Rong Jin andTieniu Tan. Free lunch for domain adversarial training: En- vironment label smoothing. InICLR, 2023

    2023

  29. [37]

    Multi- sourcemulti-modaldomainadaptation.Information Fusion, 117:102862, 2025

    Sicheng Zhao, Jing Jiang, Wenbo Tang, Jiankun Zhu, Hui Chen, Pengfei Xu, Björn W Schuller, Jianhua Tao, Hongxun Yao, and Guiguang Ding. Multi- sourcemulti-modaldomainadaptation.Information Fusion, 117:102862, 2025. Contents 1 Introduction 2 2 Related Work 4 3 Methodology 4 3.1...

    2025

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.