Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Single-Domain Generalized Object Detection by Balancing Domain Diversity and Invariance

T0 review · 4 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

Pith's one-line read This paper claims that a single-domain object detector generalizes to unseen weather and visual styles by preserving and expanding domain-specific feature diversity while aligning features only when alignment matters, and it tests this…

desk verdict A coherent diversity-plus-invariance method for single-domain detection whose own tables contradict its 'superior performance' claim, making it a borderline major-revision rather than a clear reject. read the letter →

arxiv 2502.03835 v2 pith:IPWNL545 submitted 2025-02-06 cs.CV

classification cs.CV
keywords single-domaingeneralizationobjectdetectiondomain-invariantrepresentationlearningdomain-specificfeaturediversitylossweightedalignmentdomainshifttransfer
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 is trying to establish that single-domain object detection improves when domain diversity is treated as an asset instead of a nuisance. Against the common practice of aligning everything into a domain-invariant feature space, the authors argue that a useful detector should keep the domain-specific information that survives augmentation, expand its spread, and align representations only when the alignment actually changes detection outputs. They formalize this through a target-risk bound in which the error depends both on the diameter $\rho$ of the combined source feature distribution and on its distance $\gamma$ to the unseen target, and they claim their Diversity Invariant Detection Model (DIDM) reduces both terms at once. If the claim is correct, detectors trained on a single source—clear daytime driving images, real photos, synthetic data—can generalize to unseen weather, artistic styles, and other domains without ever seeing target data.

What carries the argument

The load-bearing object is the residual feature $z_d$ defined by $z_a = \lambda_1 z_s + (1-\lambda_1)z_d$: it is the part of the feature that augmentation changes, and the paper treats it as the carrier of domain-specific diversity. Around that object, the mechanism has two coupled losses: the feature diversity loss $L_{FD}$ pushes every proposal's $z_d$ away from every other proposal's $z_d$ in cosine space under a temperature-scaled softmax, while the entropy loss $L_H$ prevents $z_d$ from encoding shared category semantics. The second mechanism is weighted alignment: $L_{WAM} = \beta(1+L(z_s,z_a))$ aligns original and augmented features, with $\beta$ dropping to nearly $1$ when classification and box outputs already agree, preventing alignment from collapsing the diversity that $L_{FD}$ builds. The paper's argument is that the first mechanism expands the source feature hull $\delta$ to reduce $\gamma$, and the second mechanism controls $\rho$, so the two losses jointly tighten the target-risk bound.

What would settle it

Train DIDM and an invariance-only baseline on the same single source, freeze a feature extractor, and compute the A-distance (or another distribution divergence) from the joint source feature hull to each held-out target domain. If increasing the feature diversity loss decreases the hull-to-target distance $\gamma$ at constant source risk, the central claim is confirmed; if $\gamma$ stays flat or grows while only the source-to-source diameter $\rho$ changes, the claimed mechanism is false.

Watch

Extended reading notes

Core claim

The central discovery the authors report is that decomposing a detector's features into an invariant part and a domain-specific part, then actively diversifying the domain-specific part, produces representations that cover unseen target domains better than invariance-only alignment. Concretely, DIDM writes the augmented feature as $z_a = \lambda_1 z_s + (1-\lambda_1)z_d$, so the residual $z_d$ carries what augmentation changed; a Diversity Learning Module then maximizes the entropy of $z_d$'s class predictions to strip category semantics, and applies a feature diversity loss that minimizes cosine similarity among all proposal-level $z_d$'s. A Weighted Aligning Module aligns $z_s$ and $z_a$ with a loss weighted by $\beta = 2 - e^{-(L_c+L_b)/2}$, which shrinks when detections on the original and augmented branches already agree, so the model does not over-align at the cost of diversity. The paper argues this joint objective tightens the bound $\xi_t(h) \le \sum_i \pi_i^\ast \xi_i(h) + (\gamma+\rho)/2 + \lambda_{H,(P_X^t,P_X^\ast)}$ by lowering both the source-hull diameter $\rho$ and the hull-to-target distance $\gamma$, and it supports the claim with experiments on weather, artistic, and synthetic-to-real detection benchmarks.

Load-bearing premise

The load-bearing premise is that pushing the domain-specific features apart makes the overall source feature distribution spread out in the directions that cover the unseen target domain, while leaving source-domain detection accuracy intact; the paper asserts this from the generalization bound without deriving it.

Editorial extensions

If this is right

  • If the mechanism works as claimed, single-domain detectors can be improved without any target-domain data simply by keeping and diversifying the features that training-time augmentation changes.
  • From clear daytime training data, the model reaches 43.4 mAP on Night-Sunny and 40.1 mAP on Daytime-Foggy, both best among the reported methods, while holding source-domain daytime-clear detection near its baseline.
  • The entropy-maximization constraint implies that the diversified residual $z_d$ is not a shortcut that memorizes category-discriminative appearance, so the detector should transfer to style shifts such as clipart, watercolor, and comic images without re-training.
  • Because the WAM alignment weight is driven by detection agreement, the model will naturally down-weight alignment in easy-condition pairs and keep aligning hard-condition pairs, which is a direct testable scheduling rule for feature alignment.

Reading between the lines

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

  • The paper's theory predicts that measured distance $\gamma$ from the learned source feature hull to each held-out target should fall as $L_{FD}$ is strengthened; a systematic sweep of $\gamma$ to every held-out target would directly test this prediction, since the reported A-distance values cover only a single source-target pair in that form.
  • Because $z_d$ is defined relative to the augmentation $\mathcal{A}$, the diversity gain is only as broad as the augmentation family; an extension would design $\mathcal{A}$ from anticipated target shifts so that the hull expands in the direction of the deployment domain.
  • The dynamic weight $\beta$ can be read as a self-paced alignment schedule: align early when outputs disagree, stop aligning once outputs agree; this suggests a curriculum interpretation in which diversity is learned first and invariance later, consistent with the increasing schedule of $\lambda_1$.
  • The decomposition is proposal-level, so the same two losses could be lifted to instance-segmentation or transformer-based detectors that produce pooled or token-level features, which would test whether the diversity mechanism is detector-agnostic.
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

4 major / 5 minor

Summary. The paper addresses single-domain generalized object detection (S-DGOD) and proposes the Diversity Invariant Detection Model (DIDM), which combines a Diversity Learning Module (DLM) and a Weighted Aligning Module (WAM). DLM is designed to preserve domain-specific feature diversity while suppressing category semantics in domain-specific features, and WAM performs feature alignment with a dynamically weighted loss. The paper motivates the method with a domain generalization bound, presents experiments on DWD, Real-to-Artistic, Diversified Weather, and Artificial-to-Real benchmarks, and claims superior performance over existing methods. The central empirical claim is that co-optimizing diversity and invariance reduces both the source-domain discrepancy and the distance to unseen target domains, yielding better generalization.

Significance. The problem of single-domain generalized object detection is practically relevant, and the idea of explicitly preserving domain-specific diversity while learning an invariant representation is reasonable. The ablation study in Table XI indicates that the proposed losses improve over the Faster R-CNN baseline on several target domains, which suggests the components are not inert. However, the paper's headline claim—superior performance over existing S-DGOD methods—is not supported by the reported comparisons against DivAlign, and the theoretical framing does not currently provide a rigorous justification for the method. If the method were shown to be consistently competitive or superior with proper uncertainty quantification, it would be a useful contribution to the S-DGOD literature.

major comments (4)
  1. [Table I and Table VII] The abstract and Section IV claim that DIDM achieves superior performance compared to existing methods, but the paper's own numbers contradict this. In Table I, the Average column (computed over the four unseen target domains, excluding the source Daytime-Clear column) is 35.1 for DIDM versus 35.5 for DivAlign; DIDM is higher on Night-Sunny and Daytime-Foggy but lower on Dusk-Rainy and Night-Rainy. In Table VII, DIDM is below DivAlign on every Real-to-Artistic target: 34.5 vs. 38.9 on Clipart1k, 31.7 vs. 33.2 on Comic2k, and 57.3 vs. 57.4 on Watercolor2k. The reported results therefore do not support the paper's central empirical claim.
  2. [Section III-A, Eqs. (2)-(5)] The theoretical motivation is based on a multi-source domain generalization bound, but the setting is single-source (M=1). For M=1, the set δ in Eq. (5) is a singleton, so the diameter ρ in Eq. (4) is identically zero and the claimed trade-off between ρ and γ is not captured by the displayed bound. The paper neither derives a single-source version of the bound nor specifies how the augmented and domain-specific features z_d enter the definition of δ. As written, Eq. (2) does not justify the paper's central mechanism.
  3. [Section III-C, IV-C, V-C] The claim that maximizing the feature diversity loss L_FD in Eq. (8) expands the joint source feature distribution δ and reduces the distance γ to any unseen target domain is asserted rather than derived. The A-distance evidence in Section V-C and Fig. 7 compares source-domain pairs (Foggy vs. Cityscapes and Rainy) and does not measure the distance to the unseen target domains used in the generalization experiments. Moreover, Eq. (2) also contains the source risk term and the ideal joint risk λ_H, and no argument is given that these terms do not degrade when L_FD is maximized.
  4. [Section IV-A and all result tables] No error bars, number of runs, seeds, or code are reported. Given that the decisive differences in Table I and Table VII are as small as 0.4 mAP and 0.1 mAP, the claim of superiority over DivAlign is not statistically established even apart from the direction of the difference.
minor comments (5)
  1. [Section III-C, Eq. (8)] The numerator in Eq. (8) is sim(f'(z_i^d), f'(z_i^d)), which equals 1 for cosine similarity; as written, the loss reduces to a contrastive-style term over the denominator only. Please clarify the intended positive pair or rewrite the expression.
  2. [Section III-C, Eq. (7)] The notation H(y|z_d) is used interchangeably with H[Softmax(f(z_d))]; the conditional entropy of a deterministic classifier output is not defined in the usual way. Please align the notation.
  3. [Section IV-C] The text states that the method 'achieves an average generalization performance of 35.0%', but Table I reports 35.1 for the Average column; the numbers should be reconciled.
  4. [Section III-E, Eq. (13)] 'WMA' in Eq. (13) should be 'WAM'.
  5. [Throughout] There are several typos and infelicities: 'data argumentation' should be 'data augmentation', 'presentation learning' should be 'representation learning', 'Cityscaps' should be 'Cityscapes', and 'argument image' should be 'augmented image'.

Circularity Check

0 steps flagged · score 1.0 of 10

No load-bearing circularity: the DG bound is motivational, the losses are explicit engineering choices evaluated on held-out targets, and the sole self-citation ([12]) is not load-bearing.

full rationale

The paper's derivation chain is not circular. The Ben-David/DG bound in Eq. (2) is imported from prior work [7], [46] and used only to motivate the two-module design; neither L_DLM in Eq. (9) nor L_WAM in Eq. (12) is obtained by solving the bound, and the claim that L_FD expands the joint distribution delta so as to reduce gamma is an asserted hypothesis, not a derived consequence. The losses themselves are explicit engineering choices: L_FD is a cosine-similarity contrastive regularizer on the domain-specific features z_d, L_H maximizes entropy of the classifier output, and the weight beta in Eq. (11) is a dynamic scalar derived from detection-result similarity. These objectives are trained on source-domain data and then evaluated on held-out target domains, so the reported numbers are genuine empirical evidence rather than restatements of the training objective. No parameter is fitted to target performance and then renamed as a prediction, and no benchmark number is constructed from the model's own assumptions. The only visible self-citation is reference [12] (Hou et al., including author Z. He), cited in the introductory sentence on domain-invariant representation learning; it is not used to justify any architecture, loss, or theorem, so it is not load-bearing. The paper does contain an unsupported leap: maximizing feature diversity is asserted, without proof, to lower the distance gamma to any unseen target domain without raising source risk. That is a rigor or correctness weakness, not circularity. Similarly, the fact that Table I and Table VII do not uniformly support the 'superior performance' headline is an empirical validity concern, not a circular-reasoning concern. Under the quoted-evidence rule, no circular step can be exhibited, so the circularity score is minimal; the single point reflects only the presence of a non-load-bearing self-citation.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The central claim rests on four free parameters (only τ and α are reported), and on two unproved modeling assumptions: that pushing z_d apart expands coverage of unseen targets, and that the interpolation cleanly separates domain-specific from invariant content. The invented z_d representation has no independent evidence of capturing what it claims to capture.

free parameters (4)
  • lambda_1 = not reported; described as gradually increasing
    Controls the interpolation z_a = λ1 z_s + (1 - λ1) z_d in Section III-C. Its schedule is unspecified, which directly affects the extracted domain-specific features.
  • lambda_2 = not reported
    Balances the feature diversity loss in Eq 9; no value or tuning protocol is given.
  • tau = 0.1
    Temperature in Eq 8; selected by hyperparameter analysis (Fig 5a) on target-domain mAP without a described held-out split.
  • alpha = 0.45
    Global loss weight in Eq 13; selected by hyperparameter analysis (Fig 5c) on target-domain mAP.
assumptions (4)
  • domain assumption The Ben-David DA bound (Eq 1) and the multi-source DG bound (Eq 2) apply to the single-source-with-augmentation setting.
    The paper uses Eq 2 as the theoretical foundation for the method, but the setting is one source domain plus augmented views, not multiple independent source domains.
  • domain assumption Source and target domains share the same label space.
    Stated in Section III-A as a problem setup condition.
  • ad hoc to paper Enhancing diversity of domain-specific features expands δ and reduces γ to unseen targets without increasing source risk.
    This is the core motivation for DLM, asserted from Eq 2 but never derived or directly measured for unseen targets.
  • ad hoc to paper The interpolation z_a = λ1 z_s + (1 - λ1) z_d gives a valid separation of invariant and domain-specific features.
    The paper defines z_d from this formula and treats it as domain-specific content, but provides no evidence that it isolates domain-specific information rather than noise.
invented entities (1)
  • z_d, the domain-specific feature representation
    purpose: Carries domain-specific information that the model preserves while aligning z_s and z_a.
    Defined only through the interpolation formula in Section III-C. No external or downstream evidence shows it isolates domain-specific content.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Single-Domain Generalized Object Detection by Balancing Domain Diversity and Invariance." pith.science (2026). https://pith.science/paper/IPWNL545

@misc{pith2026250203835,
  author       = {Pith},
  title        = {Pith review of: Single-Domain Generalized Object Detection by Balancing Domain Diversity and Invariance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IPWNL545}},
  note         = {Machine review of arXiv:2502.03835}
}
read the original abstract

Single-domain generalization for object detection (S-DGOD) seeks to transfer learned representations from a single source domain to unseen target domains. While recent approaches have primarily focused on achieving feature invariance, they ignore that domain diversity also presents significant challenges for the task. First, such invariance-driven strategies often lead to the loss of domain-specific information, resulting in incomplete feature representations. Second, cross-domain feature alignment forces the model to overlook domain-specific discrepancies, thereby increasing the complexity of the training process. To address these limitations, this paper proposes the Diversity Invariant Detection Model (DIDM), which achieves a harmonious integration of domain-specific diversity and domain invariance. Our key idea is to learn the invariant representations by keeping the inherent domain-specific features. Specifically, we introduce the Diversity Learning Module (DLM). This module limits the invariant semantics while explicitly enhancing domain-specific feature representation through a proposed feature diversity loss. Furthermore, to ensure cross-domain invariance without sacrificing diversity, we incorporate the Weighted Aligning Module (WAM) to enable feature alignment while maintaining the discriminative domain-specific information. Extensive experiments on multiple diverse datasets demonstrate the effectiveness of the proposed model, achieving superior performance compared to existing methods.

Figures

Figures reproduced from arXiv: 2502.03835 by the authors.

Figure 1
Figure 1. Illustrate three scenarios of feature distribution: (a) unaligned features, (b) fully aligned features, (c) aligned features with domain-specific information [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The general architecture of DIDM. The input image is processed through the backbone to extract both original and enhanced features, as well as to [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Qualitative evaluation results for night-sunny, night-rainy daytime-foggy and dusk-rainy weather conditions are presented. The images, arranged from [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Qualitative evaluation results for Real to Artistic datasets are presented. The [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Hyperparameter Analysis. (a) Analysis of [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Analysis of Differences in the Distribution of Features. Where Foggy represents the source domain, Cityspace and Rainy represent different target [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: A-distance Result Display. Where A represents the absence of [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Boosting Domain Generalized and Adaptive Detection with Diffusion Models: Fitness, Generalization, and Transferability

    cs.CV 2025-06 conditional novelty 5.0 of 10

    A single-step diffusion feature extractor with an object-masked auxiliary branch and consistency loss improves domain-generalized and adaptive detection accuracy and speed.

Reference graph

Works this paper leans on

71 extracted references · 58 canonical work pages · cited by 1 Pith paper

  1. [1]

    Faster r-cnn: Towards real-time object detection with region proposal networks,

    S. Ren, K. He, R. Girshick, and J. Sun, “Faster r-cnn: Towards real-time object detection with region proposal networks,” IEEE transactions on pattern analysis and machine intelligence , pp. 1137–1149, 2016

  2. [2]

    Fcos: Fully convolutional one- stage object detection. arxiv 2019,

    Z. Tian, C. Shen, H. Chen, and T. He, “Fcos: Fully convolutional one- stage object detection. arxiv 2019,” arXiv preprint arXiv:1904.01355 , 2019

  3. [3]

    Yolox: Exceeding yolo series in 2021,

    Z. Ge, “Yolox: Exceeding yolo series in 2021,” arXiv preprint arXiv:2107.08430, 2021

  4. [4]

    Generalized and discriminative few-shot object detection via svd-dictionary enhancement,

    A. Wu, S. Zhao, C. Deng, and W. Liu, “Generalized and discriminative few-shot object detection via svd-dictionary enhancement,” Advances in Neural Information Processing Systems , pp. 6353–6364, 2021

  5. [5]

    Harmonizing transferability and discriminability for adapting object detectors,

    C. Chen, Z. Zheng, X. Ding, Y . Huang, and Q. Dou, “Harmonizing transferability and discriminability for adapting object detectors,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 8869–8878

  6. [6]

    Unbiased look at dataset bias,

    A. Torralba and A. A. Efros, “Unbiased look at dataset bias,” in CVPR

  7. [7]

    Generalizing to unseen domains: A survey on domain generalization,

    J. Wang, C. Lan, C. Liu, Y . Ouyang, T. Qin, W. Lu, Y . Chen, W. Zeng, and S. Y . Philip, “Generalizing to unseen domains: A survey on domain generalization,” IEEE transactions on knowledge and data engineering , vol. 35, no. 8, pp. 8052–8072, 2022

  8. [8]

    A theory of learning from different domains,

    S. Ben-David, J. Blitzer, K. Crammer, A. Kulesza, F. Pereira, and J. W. Vaughan, “A theory of learning from different domains,” Machine learning, vol. 79, pp. 151–175, 2010

Show all 71 references
  1. [9]

    Domain-adversarial training of neural networks,

    Y . Ganin, E. Ustinova, H. Ajakan, P. Germain, H. Larochelle, F. Lavi- olette, M. March, and V . Lempitsky, “Domain-adversarial training of neural networks,” Journal of machine learning research, vol. 17, no. 59, pp. 1–35, 2016

  2. [10]

    Optimal transport for domain adaptation,

    N. Courty, R. Flamary, D. Tuia, and A. Rakotomamonjy, “Optimal transport for domain adaptation,” IEEE transactions on pattern analysis and machine intelligence , vol. 39, no. 9, pp. 1853–1865, 2016

  3. [11]

    Domain invariant representation learning with domain density transformations,

    A. T. Nguyen, T. Tran, Y . Gal, and A. G. Baydin, “Domain invariant representation learning with domain density transformations,” Advances in Neural Information Processing Systems , vol. 34, pp. 5264–5275, 2021

  4. [12]

    Gradient-aware domain-invariant learning for domain generalization,

    F. Hou, Y . Zhang, Y . Liu, J. Yuan, C. Zhong, Y . Zhang, Z. Shi, J. Fan, and Z. He, “Gradient-aware domain-invariant learning for domain generalization,” Multimedia Systems, vol. 31, no. 1, pp. 1–15, 2025

  5. [13]

    Do- main generalization via model-agnostic learning of semantic features,

    Q. Dou, D. Coelho de Castro, K. Kamnitsas, and B. Glocker, “Do- main generalization via model-agnostic learning of semantic features,” Advances in neural information processing systems , vol. 32, 2019

  6. [14]

    Unbiased faster r-cnn for single-source domain generalized object detection,

    Y . Liu, S. Zhou, X. Liu, C. Hao, B. Fan, and J. Tian, “Unbiased faster r-cnn for single-source domain generalized object detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 28 838–28 847

  7. [15]

    Prompt-driven dynamic object- centric learning for single domain generalization,

    D. Li, A. Wu, Y . Wang, and Y . Han, “Prompt-driven dynamic object- centric learning for single domain generalization,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 17 606–17 615

  8. [16]

    G-nas: Generalizable neural architecture search for single domain generalization object detection,

    F. Wu, J. Gao, L. Hong, X. Wang, C. Zhou, and N. Ye, “G-nas: Generalizable neural architecture search for single domain generalization object detection,” in Proceedings of the AAAI Conference on Artificial Intelligence, 2024, pp. 5958–5966

  9. [17]

    Domain general- ization: A survey,

    K. Zhou, Z. Liu, Y . Qiao, T. Xiang, and C. C. Loy, “Domain general- ization: A survey,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 4, pp. 4396–4415, 2022

  10. [18]

    Wildnet: Learning domain generalized semantic segmentation from the wild,

    S. Lee, H. Seong, S. Lee, and E. Kim, “Wildnet: Learning domain generalized semantic segmentation from the wild,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 9936–9946

  11. [19]

    Deep discriminative causal domain generalization,

    S. Li, Q. Zhao, C. Zhang, and Y . Zou, “Deep discriminative causal domain generalization,” Information Sciences, vol. 645, p. 119335, 2023

  12. [20]

    Frustratingly simple domain generalization via image stylization,

    N. Somavarapu, C.-Y . Ma, and Z. Kira, “Frustratingly simple domain generalization via image stylization,” arXiv preprint arXiv:2006.11207 , 2020

  13. [21]

    Learning to diversify for single domain generalization,

    Z. Wang, Y . Luo, R. Qiu, Z. Huang, and M. Baktashmotlagh, “Learning to diversify for single domain generalization,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2021, pp. 834–843

  14. [22]

    Multi-adversarial discriminative deep domain generalization for face presentation attack detection,

    R. Shao, X. Lan, J. Li, and P. C. Yuen, “Multi-adversarial discriminative deep domain generalization for face presentation attack detection,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 10 023–10 031

  15. [23]

    Meta-causal learning for single domain generalization,

    J. Chen, Z. Gao, X. Wu, and J. Luo, “Meta-causal learning for single domain generalization,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023, pp. 7683–7692

  16. [24]

    Learning to learn with variational information bottleneck for domain generalization,

    Y . Du, J. Xu, H. Xiong, Q. Qiu, X. Zhen, C. G. Snoek, and L. Shao, “Learning to learn with variational information bottleneck for domain generalization,” in Computer Vision–ECCV 2020: 16th European Con- ference, Glasgow, UK, August 23–28, 2020, Proceedings, Part X 16 . Spring...

  17. [25]

    Semantic-aware domain generalized segmentation,

    D. Peng, Y . Lei, M. Hayat, Y . Guo, and W. Li, “Semantic-aware domain generalized segmentation,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 2594–2605

  18. [26]

    Learning to optimize domain specific normalization for domain generalization,

    S. Seo, Y . Suh, D. Kim, G. Kim, J. Han, and B. Han, “Learning to optimize domain specific normalization for domain generalization,” in Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXII 16 . Springer, 2020, pp. 68–83

  19. [27]

    Improving generalization of meta-learning with inverted regularization at inner- level,

    L. Wang, S. Zhou, S. Zhang, X. Chu, H. Chang, and W. Zhu, “Improving generalization of meta-learning with inverted regularization at inner- level,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 7826–7835

  20. [28]

    Learning to generalize unseen domains via memory-based multi-source meta- learning for person re-identification,

    Y . Zhao, Z. Zhong, F. Yang, Z. Luo, Y . Lin, S. Li, and N. Sebe, “Learning to generalize unseen domains via memory-based multi-source meta- learning for person re-identification,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. ...

  21. [29]

    I3net: Implicit instance-invariant network for adapting one-stage object detectors,

    C. Chen, Z. Zheng, Y . Huang, X. Ding, and Y . Yu, “I3net: Implicit instance-invariant network for adapting one-stage object detectors,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 12 576–12 585

  22. [30]

    Unbiased mean teacher for cross- domain object detection,

    J. Deng, W. Li, Y . Chen, and L. Duan, “Unbiased mean teacher for cross- domain object detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2021, pp. 4091–4101

  23. [31]

    Mila: memory-based instance- level adaptation for cross-domain object detection,

    O. Krishna, H. Ohashi, and S. Sinha, “Mila: memory-based instance- level adaptation for cross-domain object detection,” arXiv preprint arXiv:2309.01086, 2023

  24. [32]

    Cross-domain adaptive teacher for object detection,

    Y .-J. Li, X. Dai, C.-Y . Ma, Y .-C. Liu, K. Chen, B. Wu, Z. He, K. Kitani, and P. Vajda, “Cross-domain adaptive teacher for object detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 7581–7590

  25. [33]

    Unsupervised domain adaptation of object detectors: A survey,

    P. Oza, V . A. Sindagi, V . V . Sharmini, and V . M. Patel, “Unsupervised domain adaptation of object detectors: A survey,” IEEE Transactions on Pattern Analysis and Machine Intelligence , 2023

  26. [34]

    A review of single-source deep unsupervised visual domain adaptation,

    S. Zhao, X. Yue, S. Zhang, B. Li, H. Zhao, B. Wu, R. Krishna, J. E. Gonzalez, A. L. Sangiovanni-Vincentelli, S. A. Seshia et al., “A review of single-source deep unsupervised visual domain adaptation,” IEEE Transactions on Neural Networks and Learning Systems , vol. 33, no. 2,...

  27. [35]

    Contrastive mean teacher for domain adaptive object detectors,

    S. Cao, D. Joshi, L.-Y . Gui, and Y .-X. Wang, “Contrastive mean teacher for domain adaptive object detectors,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 23 839–23 848

  28. [36]

    Strong-weak distri- bution alignment for adaptive object detection,

    K. Saito, Y . Ushiku, T. Harada, and K. Saenko, “Strong-weak distri- bution alignment for adaptive object detection,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 6956–6965

  29. [37]

    Unified domain generalization and adaptation for multi-view 3d object detection,

    G. Chang, J. Lee, D. Kim, J. Kim, D. Lee, D. Ji, S. Jang, and S. Kim, “Unified domain generalization and adaptation for multi-view 3d object detection,” arXiv preprint arXiv:2410.22461 , 2024

  30. [38]

    Towards generalizable multi-object tracking,

    Z. Qin, L. Wang, S. Zhou, P. Fu, G. Hua, and W. Tang, “Towards generalizable multi-object tracking,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 18 995–19 004

  31. [39]

    Clip the gap: A single domain generalization approach for object detection,

    V . Vidit, M. Engilberge, and M. Salzmann, “Clip the gap: A single domain generalization approach for object detection,” in Proceedings of JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 14 the IEEE/CVF conference on computer vision and pattern recognition , 2023, pp...

  32. [40]

    Improving single domain-generalized object detection: A focus on diversification and alignment,

    M. S. Danish, M. H. Khan, M. A. Munir, M. S. Sarfraz, and M. Ali, “Improving single domain-generalized object detection: A focus on diversification and alignment,” in Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, 2024, pp. 17 732– 17 742

  33. [41]

    Object-aware domain gen- eralization for object detection,

    W. Lee, D. Hong, H. Lim, and H. Myung, “Object-aware domain gen- eralization for object detection,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 38, no. 4, 2024, pp. 2947–2955

  34. [42]

    Two at once: Enhancing learning and generalization capacities via ibn-net,

    X. Pan, P. Luo, J. Shi, and X. Tang, “Two at once: Enhancing learning and generalization capacities via ibn-net,” in Proceedings of the euro- pean conference on computer vision (ECCV) , 2018, pp. 464–479

  35. [43]

    Robustnet: Improving domain generalization in urban-scene segmentation via in- stance selective whitening,

    S. Choi, S. Jung, H. Yun, J. T. Kim, S. Kim, and J. Choo, “Robustnet: Improving domain generalization in urban-scene segmentation via in- stance selective whitening,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2021, pp. 11 580–11 590

  36. [44]

    Iterative normalization: Beyond standardization towards efficient whitening,

    L. Huang, Y . Zhou, F. Zhu, L. Liu, and L. Shao, “Iterative normalization: Beyond standardization towards efficient whitening,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2019, pp. 4874–4883

  37. [45]

    Single-domain generalized object detection in urban scene via cyclic-disentangled self-distillation,

    A. Wu and C. Deng, “Single-domain generalized object detection in urban scene via cyclic-disentangled self-distillation,” in Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition , 2022, pp. 847–856

  38. [46]

    Adversar- ial target-invariant representation learning for domain generalization,

    I. Albuquerque, J. Monteiro, T. H. Falk, and I. Mitliagkas, “Adversar- ial target-invariant representation learning for domain generalization,” CoRR, 2019

  39. [47]

    Order-preserving consistency regularization for domain adaptation and generalization,

    M. Jing, X. Zhen, J. Li, and C. G. Snoek, “Order-preserving consistency regularization for domain adaptation and generalization,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2023, pp. 18 916–18 927

  40. [48]

    Switchable whitening for deep representation learning,

    X. Pan, X. Zhan, J. Shi, X. Tang, and P. Luo, “Switchable whitening for deep representation learning,” in Proceedings of the IEEE/CVF international conference on computer vision , 2019, pp. 1863–1871

  41. [49]

    Style-hallucinated dual consistency learning for domain generalized semantic segmenta- tion,

    Y . Zhao, Z. Zhong, N. Zhao, N. Sebe, and G. H. Lee, “Style-hallucinated dual consistency learning for domain generalized semantic segmenta- tion,” in European conference on computer vision . Springer, 2022, pp. 535–552

  42. [50]

    Srcd: Se- mantic reasoning with compound domains for single-domain generalized object detection,

    Z. Rao, J. Guo, L. Tang, Y . Huang, X. Ding, and S. Guo, “Srcd: Se- mantic reasoning with compound domains for single-domain generalized object detection,” IEEE Transactions on Neural Networks and Learning Systems, 2024

  43. [51]

    Sdg-yolov8: Single-domain generalized object detection based on domain diversity in traffic road scenes,

    H. Wang and H. Qian, “Sdg-yolov8: Single-domain generalized object detection based on domain diversity in traffic road scenes,” Displays, vol. 87, p. 102948, 2025

  44. [52]

    Let synthetic data shine: Domain reassembly and soft-fusion for single domain generalization,

    H. Li, Y . Xiao, K. Liang, M. Wang, L. Lan, K. Li, and X. Liu, “Let synthetic data shine: Domain reassembly and soft-fusion for single domain generalization,” arXiv preprint arXiv:2503.13617 , 2025

  45. [53]

    An enhanced domain generalization method for object detection based on text guided feature disentangle- ment,

    M. Wang, Y . Liu, and H. Liu, “An enhanced domain generalization method for object detection based on text guided feature disentangle- ment,” Digital Signal Processing , vol. 156, p. 104855, 2025

  46. [54]

    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

  47. [55]

    Bdd100k: A diverse driving dataset for heterogeneous multitask learning,

    F. Yu, H. Chen, X. Wang, W. Xian, Y . Chen, F. Liu, V . Madhavan, and T. Darrell, “Bdd100k: A diverse driving dataset for heterogeneous multitask learning,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2020, pp. 2636–2645

  48. [56]

    Vehicle detection and tracking in adverse weather using a deep learning frame- work,

    M. Hassaballah, M. A. Kenk, K. Muhammad, and S. Minaee, “Vehicle detection and tracking in adverse weather using a deep learning frame- work,” IEEE transactions on intelligent transportation systems , vol. 22, no. 7, pp. 4230–4242, 2020

  49. [57]

    The cityscapes dataset for semantic urban scene understanding,

    M. Cordts, M. Omran, S. Ramos, T. Rehfeld, M. Enzweiler, R. Be- nenson, U. Franke, S. Roth, and B. Schiele, “The cityscapes dataset for semantic urban scene understanding,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 3213– 3223

  50. [58]

    Vector-decomposed disentanglement for domain-invariant object detection,

    A. Wu, R. Liu, Y . Han, L. Zhu, and Y . Yang, “Vector-decomposed disentanglement for domain-invariant object detection,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2021, pp. 9342–9351

  51. [59]

    The pascal visual object classes (voc) challenge,

    M. Everingham, L. Van Gool, C. K. Williams, J. Winn, and A. Zisser- man, “The pascal visual object classes (voc) challenge,” International journal of computer vision , vol. 88, pp. 303–338, 2010

  52. [60]

    Cross-domain weakly-supervised object detection through progressive domain adap- tation,

    N. Inoue, R. Furuta, T. Yamasaki, and K. Aizawa, “Cross-domain weakly-supervised object detection through progressive domain adap- tation,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 5001–5009

  53. [61]

    Semantic foggy scene under- standing with synthetic data,

    C. Sakaridis, D. Dai, and L. Van Gool, “Semantic foggy scene under- standing with synthetic data,” International Journal of Computer Vision, vol. 126, pp. 973–992, 2018

  54. [62]

    Depth-attentional features for single-image rain removal,

    X. Hu, C.-W. Fu, L. Zhu, and P.-A. Heng, “Depth-attentional features for single-image rain removal,” in Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition , 2019, pp. 8022–8031

  55. [63]

    Driving in the matrix: Can virtual worlds replace human-generated annotations for real world tasks?

    M. Johnson-Roberson, C. Barto, R. Mehta, S. N. Sridhar, and R. Vasude- van, “Driving in the matrix: Can virtual worlds replace human-generated annotations for real world tasks?” in IEEE International Conference on Robotics & Automation , 2017

  56. [64]

    Dual bipartite graph learning: A general approach for domain adaptive object detection,

    C. Chen, J. Li, Z. Zheng, Y . Huang, X. Ding, and Y . Yu, “Dual bipartite graph learning: A general approach for domain adaptive object detection,” in Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 2703–2712

  57. [65]

    Towards robust object detection invariant to real-world domain shifts,

    Q. Fan, M. Segu, Y .-W. Tai, F. Yu, C.-K. Tang, B. Schiele, and D. Dai, “Towards robust object detection invariant to real-world domain shifts,” in The Eleventh International Conference on Learning Representations (ICLR 2023). OpenReview, 2023

  58. [66]

    Generalized diffusion detector: Mining robust features from diffusion models for domain- generalized detection,

    B. He, Y . Ji, Q. Ye, Z. Tan, and L. Wu, “Generalized diffusion detector: Mining robust features from diffusion models for domain- generalized detection,” in Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 9921–9932

  59. [67]

    Boosting domain generalized and adaptive detection with diffusion models: Fitness, generalization, and transferability,

    B. He, Y . Ji, Z. Tan, and L. Wu, “Boosting domain generalized and adaptive detection with diffusion models: Fitness, generalization, and transferability,” arXiv preprint arXiv:2506.21042 , 2025

  60. [68]

    A fourier-based framework for domain generalization,

    Q. Xu, R. Zhang, Y . Zhang, Y . Wang, and Q. Tian, “A fourier-based framework for domain generalization,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2021, pp. 14 383–14 392

  61. [69]

    Frequency space domain randomization for domain generalization,

    J. Huang, D. Guan, A. Xiao, and S. L. FSDR, “Frequency space domain randomization for domain generalization,” in Proc. of CVPR , 2021, pp. 6891–6902

  62. [70]

    Multi-view adversarial discriminator: Mine the non-causal factors for object detection in unseen domains,

    M. Xu, L. Qin, W. Chen, S. Pu, and L. Zhang, “Multi-view adversarial discriminator: Mine the non-causal factors for object detection in unseen domains,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2023, pp. 8103–8112

  63. [2011]

    1521–1528

    IEEE, 2011, pp. 1521–1528

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.