Pith. sign in

REVIEW 3 major objections 5 minor 81 references

UPRE: Zero-Shot Domain Adaptation for Object Detection via Unified Prompt and Representation Enhancement

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that jointly optimizing learnable prompts and visual feature enhancers lets a detector adapt to unseen target domains without ever seeing a target image.

desk verdict A plausible incremental ZSDA detection method whose central claim is undercut by mAP/per-class inconsistencies and an unvalidated pseudo-target mechanism. read the letter →

arxiv 2507.00721 v2 pith:URRVXRGX submitted 2025-07-01 cs.CV

Xiao Zhang , Fei Wei , Yong Wang , Wenda Zhao , Feiyi Li , Xiangxiang Chu This is my paper · ORCID
classification cs.CV
keywords zero-shotdomainadaptationobjectdetectionvision-languagemodelspromptlearningrepresentationenhancementgeneralizationbiasmulti-modalalignment
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

Zero-shot domain adaptation for object detection must prepare a detector for a target domain it has never seen, whether that means different weather, a different city, or a synthetic-to-real shift. Existing vision-language approaches use prompts to fight the distribution shift, but they leave a second problem untreated: the detection task itself is misaligned with a model trained to describe whole images. UPRE trains a multi-view learnable domain prompt and a visual representation enhancer in one loop, so the text side and the image side refine each other. On nine benchmark datasets the paper reports consistent mAP gains over prior methods, including 40.0 versus 39.6 on Daytime Foggy and 74.3 versus 73.6 AP for the Cityscapes-to-KITTI car. The claim is that this unified treatment beats spending the zero-shot budget on prompts alone or on style transfer alone.

What carries the argument

The load-bearing mechanism is the patchwise affine style transform of Eq. (5): source feature maps are split into $M\times N$ patches and each patch $j$ is shifted and scaled by learnable tensors, $F_{s\to t}=\{E^j_\sigma \cdot F^j_s + E^j_\mu\}_{j=1}^{M\times N}$, to synthesize pseudo-target features without any target image. MDP supplies the language side, retaining the static prompt so the learnable context can concentrate on domain-adaptation knowledge for image, object, and background views. RDD aligns image-level embeddings with prompt embeddings while an $L_1$ constraint preserves source semantics and a relative-distance term stabilizes the conflicting enhancement and constraint objectives. PNS separates positive and negative proposals, giving the negative prompt its own reciprocal-label loss over $C\cup C_{\mathrm{bg}}$ so background context is learned rather than collapsed. The design point is that the prompt and the enhancer co-train: prompt updates redirect the pseudo-target style, and richer pseudo-target features feed detection knowledge back into the prompt.

What would settle it

Hold the full UPRE pipeline fixed but replace the target prompt with a deliberately wrong one, such as using "night" when testing on foggy daytime, and compare mAP on Daytime Foggy against the $L_a$-only ablation; if the gain persists, the improvement is regularization rather than genuine domain alignment. Additionally, train a domain classifier to separate $F_{s\to t}$ from real held-out target features, or compute a distribution distance such as FID; if pseudo-target and real target features are easily separable, Eq. (5) is not modeling the target domain and the alignment losses are working on a fiction.

Watch

Extended reading notes

Core claim

The central claim is that detection bias and domain bias in vision-language zero-shot domain adaptation can be handled together instead of traded off. The paper proposes a multi-view domain prompt (MDP) that keeps a static human-written domain phrase and adds learnable context vectors for three views: image-level, positive foreground proposals, and negative background proposals. Simultaneously, a unified representation enhancement (URE) module manufactures pseudo-target features from source features with a per-patch affine transform $F_{s\to t}=E_\sigma\cdot F_s+E_\mu$, where $E_\mu$ and $E_\sigma$ are learnable mean and deviation tensors. Two strategies align these synthetic features with the prompts: relative domain distance (RDD) acts at image level by matching relative distances between image embeddings to distances between text embeddings, and positive-negative separation (PNS) acts at instance level by training foreground and background proposal losses separately. After this stage the prompt, enhancer, and text encoder are frozen while the CLIP backbone and Faster R-CNN detector are fine-tuned with pseudo-target features applied half the time. The paper's discovery is the method-level result: on nine datasets spanning adverse weather, cross-city, and virtual-to-real shifts, this joint training records the best mAP, e.g. 40.0 on Daytime Foggy, 41.5 on Night Clear, and 74.3 AP for the Cityscapes-to-KITTI car.

Load-bearing premise

The method assumes that the synthetic target-style features it manufactures from source images are close enough to the real target domain's feature distribution that aligning to them helps the detector, even though the affine transform parameters are learned only from source data and text descriptions, and no target image is ever seen.

Editorial extensions

If this is right

  • Zero-shot domain adaptation can be run as a single joint objective instead of a prompt-tuning pass followed by a separate detector fine-tune.
  • Region-level style synthesis should generalize better than global style transfer to realistic scenes whose weather and lighting vary across the image.
  • Explicitly training the negative prompt against the background category should improve detection in cluttered scenes where proposals mix object and context.
  • The same frozen-at-inference design means the method adds no target-domain runtime cost beyond the fine-tuned detector.
  • The cross-city and virtual-to-real results imply the mechanism is not weather-specific and transfers to geographic and synthetic-to-real gaps.

Reading between the lines

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

  • Because the paper never directly compares pseudo-target features with real target features, a natural extension is to test whether the per-patch transform actually matches the target distribution; if it mostly acts as a strong regularizer, the relative-domain-distance recipe could be simplified.
  • The relative-distance formulation is portable: any dense prediction task that aligns image and text embeddings could reuse the idea of matching relative image-embedding shifts to relative text-embedding shifts.
  • The largest per-class gains land on hard categories such as bus, motor, and rider, suggesting that a class-weighted or difficulty-aware version of positive-negative separation might push further.
  • Frozen text encoder and early image blocks leave a clear headroom question: unfreezing more of the backbone may change the balance between prompt and enhancement, so the reported gains should be rechecked at higher compute budgets.
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 UPRE, a zero-shot domain adaptation (ZSDA) framework for object detection that jointly optimizes a multi-view domain prompt (MDP) and a unified representation enhancement (URE) module. URE generates pseudo-target-domain features from source features via per-patch affine transformations (Eq. 5), and the framework adds image-level relative domain distance (RDD) and instance-level positive-negative separation (PNS) losses. Training is performed only on source data; inference disables URE. The method is evaluated on nine datasets across three settings (adverse weather, cross-city, virtual-to-real), with per-class AP50 and mAP tables and ablations, and the paper claims superior performance over prior ZSDA and single-domain generalization methods.

Significance. If the central mechanism is supported, UPRE is a useful contribution: it addresses both domain bias and detection bias in a single framework, and the evaluation is broad (nine datasets, three domain-shift settings, component ablations). The paper also releases code and includes useful ablations of the prompt design, the enhancement granularity, and the PNS loss. However, the current evidence for the claimed mechanism is incomplete, and the quantitative tables contain internal inconsistencies that must be resolved before the performance claims can be accepted.

major comments (3)
  1. [Sec. 3.2, Eq. (5); Sec. 3.3, Eqs. (6)-(12); Sec. 3.4] The central mechanism is not verified: the per-patch affine parameters E_mu and E_sigma are learned only on source features and are supervised indirectly through alignment to CLIP text embeddings, not through any comparison with real target-domain features. No experiment establishes that F_{s->t} is actually closer to the target feature manifold than F_s is, or that the learned style parameters are semantically meaningful. Because the reported gains over prior methods are small (mostly 0.4-1.7 mAP), the alternative explanation that the enhancement acts as a stochastic regularizer rather than as domain adaptation is not excluded. The point is not circularity, since evaluation is on real target images, but the mechanism support is missing. Please add a direct diagnostic, for example comparing feature statistics of F_{s->t} with real target-domain features from the same embeddings used in Fig. 5, or ablate the learned enhancement against fixed/random augmentations in the identical training pipeline.
  2. [Tables 2-5] The reported mAP values are inconsistent with the per-class AP50 numbers in the same tables. In Table 3 the seven per-class AP50 values average to 20.5, not the reported 19.8; in Table 4 they average to 35.3, not 34.5; in Table 5 they average to 40.7, not 41.5; and in Table 2 they average to 39.7, not 40.0. If mAP is not the unweighted mean of the listed AP50 values, its definition must be stated explicitly; otherwise the numbers must be corrected. This is load-bearing because the paper's central claim of superior performance across nine benchmarks rests on these mAP values.
  3. [Sec. 4.3, Tables 1, 3, 6, 7] The reported improvements over prior methods are small, with the largest being 1.7 mAP on BDD100K in Table 6 and most below 1 mAP. No standard deviations, number of runs, or significance tests are reported. Under standard stochastic object-detection training, differences of this size are not interpretable without variance information. Please report mean and standard deviation over at least three runs, or otherwise justify why the training is deterministic and single-run comparisons are meaningful.
minor comments (5)
  1. [Sec. 3.2, Eq. (5)] The notation for the per-patch enhancement is ambiguous: please specify how the M x N patch index j is broadcast to the feature tensor and how the learnable tensors E_mu and E_sigma interact with the channel dimension C.
  2. [Sec. 3.3, Eq. (12)] The text says 'only considering positive differences' but Eq. (12) is a standard soft-label cross-entropy with label 1/(C+C_bg); please clarify the relationship between the described mechanism and the equation.
  3. [Fig. 4 caption] The word 'donate' should be 'denote' in the caption.
  4. [Sec. 2.2] There is a typo: 'doamin-invariant' should be 'domain-invariant'.
  5. [Fig. 2 caption] The sentence 'The symbols represent losses' appears to be missing the actual symbol references; please complete the caption.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation: UPRE's central claim is evaluated on real target-domain images, not on its internally generated pseudo-target features or fitted losses.

full rationale

The paper's claimed derivation chain is empirical rather than definitional. The pseudo-target features F_s->t produced by Eq. (5) are a training-time construct, supervised by text-alignment losses (Eqs. 6-8) and instance-level losses (Eqs. 9-12), but the reported results are mAP values on real, unseen target-domain images under the standard ZSDA protocol. The learned enhancement parameters E_mu and E_sigma are not fitted to the evaluation metric or to target-domain features, so no fitted input is renamed as a prediction. The joint optimization of prompts and representations is an internal training objective whose effectiveness is tested through external benchmarks and ablations, not assumed by construction. The self-citations to the authors' prior work appear in unrelated contexts (VLM backbone lists and efficiency comparisons) and do not carry the domain-adaptation argument. Although the target-domain linguistic descriptions in the prompt templates could be debated as a form of privileged information, they are part of the ZSDA setup and do not make the evaluation metric equal to a training loss. The small margins over prior methods are a correctness or significance concern, not circularity.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on CLIP's cross-modal alignment and on the affine style-transform assumption. Learnable prompts and style parameters are trained on source data; these are part of the method, not independent predictions. The loss weights and patch size are unspecified hyperparameters.

free parameters (5)
  • learnable prompt context vectors (u_l, v_l, w_l) = learned during Stage 1
    Introduced in Eqs. (2)-(4) as learnable context tokens that carry domain-adaptation and detection knowledge. Their values are optimized on source domain; no independent evidence for their necessity beyond the ablation.
  • patch-level style parameters E_mu, E_sigma = learned during Stage 1
    Introduced in Eq. (5) as C x M x N tensors that modify mean and deviation per patch to synthesize pseudo target features.
  • loss weights for L_a, L_s, L_r, L_c, L_bg = not reported
    The paper gives no weights for the five losses; presumably set to 1.0. This is a tuning choice that could affect results.
  • patch grid size M x N = 7 x 7 (diverse weather)
    M and N are set to 7x7 to match the CLIP attention pool feature size; this is a design choice that could be dataset-specific.
  • enhancement probability during fine-tuning = 0.5
    Random application of URE to source features with probability 50%, following CLIP-GAP [51].
assumptions (4)
  • domain assumption CLIP text embeddings for domain descriptions (e.g., 'foggy day') are aligned with the visual feature distribution of that domain.
    The entire method relies on this to generate pseudo target features and to guide RDD and prompt learning (Sections 3.1 and 3.3).
  • domain assumption The target domain visual shift can be approximated by a per-patch affine transformation of source features with learned mean and deviation.
    URE Eq. (5) assumes this transformation yields pseudo target features that are both style-diverse and semantically faithful.
  • domain assumption Relative distances in CLIP embedding space are meaningful: the vector difference between source and target text embeddings corresponds to the visual domain shift.
    L_r in Eq. (8) directly uses this to regularize the visual shift.
  • ad hoc to paper Background proposals should be equally dissimilar to all classes including background, with label 1/(C+C_bg).
    PNS Eq. (12) uses this soft label to train the negative prompt; no evidence is given that this is optimal.

how reviews work

0 comments
Cite this review

Pith. "Pith review of UPRE: Zero-Shot Domain Adaptation for Object Detection via Unified Prompt and Representation Enhancement." pith.science (2026). https://pith.science/paper/URRVXRGX

@misc{pith2026250700721,
  author       = {Pith},
  title        = {Pith review of: UPRE: Zero-Shot Domain Adaptation for Object Detection via Unified Prompt and Representation Enhancement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/URRVXRGX}},
  note         = {Machine review of arXiv:2507.00721}
}
read the original abstract

Zero-shot domain adaptation (ZSDA) presents substantial challenges due to the lack of images in the target domain. Previous approaches leverage Vision-Language Models (VLMs) to tackle this challenge, exploiting their zero-shot learning capabilities. However, these methods primarily address domain distribution shifts and overlook the misalignment between the detection task and VLMs, which rely on manually crafted prompts. To overcome these limitations, we propose the unified prompt and representation enhancement (UPRE) framework, which jointly optimizes both textual prompts and visual representations. Specifically, our approach introduces a multi-view domain prompt that combines linguistic domain priors with detection-specific knowledge, and a visual representation enhancement module that produces domain style variations. Furthermore, we introduce multi-level enhancement strategies, including relative domain distance and positive-negative separation, which align multi-modal representations at the image level and capture diverse visual representations at the instance level, respectively. Extensive experiments conducted on nine benchmark datasets demonstrate the superior performance of our framework in ZSDA detection scenarios. Code is available at https://github.com/AMAP-ML/UPRE.

Figures

Figures reproduced from arXiv: 2507.00721 by the authors.

Figure 1
Figure 1. (a) Prompt-driven semantic augmentation generates pseudo target domain features with manually crafted prompts. (b) The proposed unified prompt and representation enhancement creates the synthetic target domain representations via learnable prompts. Learnable prompt and unified enhancement jointly miti￾gate detection and domain biases. IE represents the image encoder, and TE denotes the text encoder. has markedly adv… view at source ↗
Figure 2
Figure 2. Overview of unified prompt and representation enhancement. Initially, the multi-view domain prompt provides language prior while also learns adaptation knowledge for object detection. Meanwhile, unified representation enhancement enriches source domain features with RPN, generating pseudo target domain representations. Finally, two enhancement strategies constrain language and visual representations, training the un… view at source ↗
Figure 3
Figure 3. Illustration of Our Proposed Strategies. Contradictory losses represent L𝑎 and L𝑠. The symbols and denote the rep￾resentations of the source domain, pseudo target domain. and represent the representations of image prompts for the source and target domains. Additionally, the cross-entropy loss L𝑐𝑒 assigns proposals to all possible foregrounds. The symbols , , , and indicate the representations of negative prompt, pos… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Qualitative Results on Diverse Weather Conditions. The top and bottom rows donate the results of OA-DG [29] and our proposed UPRE, respectively. Method Bus Bike Car Motor Person Rider Truck mAP Faster-RCNN[48] 30.7 26.7 49.7 26.2 30.9 35.5 23.2 32.0 S-DGOD[60] 32.9 28.…
Figure 5
Figure 5. Figure 5: The t-SNE visualization of image embeddings across five [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: The domain prompt templates used for zero-shot domain adaptation object detection. [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 1
Figure 1. Figure 1: We apply L2 normalization to obtain the final [PITH_FULL_IMAGE:figures/full_fig_p014_1.png]
Figure 7
Figure 7. Figure 7: Visualization results under cross-city scenarios. The top to bottom rows show results from CLIP-GAP [ [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: Visualization results of virtual-to-real world transitions. The top to bottom rows show results from CLIP-GAP [ [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

81 extracted references · 70 canonical work pages

  1. [1]

    Flamingo: a visual language model for few-shot learning

    Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716–23736,

  2. [2]

    Zodi: Zero- shot domain adaptation with diffusion-based image transfer

    Hiroki Azuma, Yusuke Matsui, and Atsuto Maki. Zodi: Zero- shot domain adaptation with diffusion-based image transfer. arXiv preprint arXiv:2403.13652, 2024. 1

    arXiv 2024

  3. [3]

    Diprompt: Disentan- gled prompt tuning for multiple latent domain generalization in federated learning

    Sikai Bai, Jie Zhang, Song Guo, Shuaicheng Li, Jingcai Guo, Jun Hou, Tao Han, and Xiaocheng Lu. Diprompt: Disentan- gled prompt tuning for multiple latent domain generalization in federated learning. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 27284–27293, 2024. 2

    work page 2024

  4. [4]

    Multi-source domain adaptation for object detection with prototype-based mean teacher

    Atif Belal, Akhil Meethal, Francisco Perdigon Romero, Marco Pedersoli, and Eric Granger. Multi-source domain adaptation for object detection with prototype-based mean teacher. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1277–1286, 2024. 1

    work page 2024

  5. [5]

    Stylip: Multi-scale style- conditioned prompt learning for clip-based domain general- ization

    Shirsha Bose, Ankit Jha, Enrico Fini, Mainak Singha, Elisa Ricci, and Biplab Banerjee. Stylip: Multi-scale style- conditioned prompt learning for clip-based domain general- ization. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5542–5552, 2024. 2

    work page 2024

  6. [6]

    Domain-controlled prompt learning

    Qinglong Cao, Zhengqin Xu, Yuntian Chen, Chao Ma, and Xiaokang Yang. Domain-controlled prompt learning. InPro- ceedings of the AAAI Conference on Artificial Intelligence , volume 38, pages 936–944, 2024. 2, 4

    work page 2024

  7. [7]

    Pali: A jointly- scaled multilingual language-image model

    Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, et al. Pali: A jointly- scaled multilingual language-image model. arXiv preprint arXiv:2209.06794, 2022. 1

    arXiv 2022

  8. [8]

    Disentangled prompt rep- resentation for domain generalization

    De Cheng, Zhipeng Xu, Xinyang Jiang, Nannan Wang, Dongsheng Li, and Xinbo Gao. Disentangled prompt rep- resentation for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23595–23604, 2024. 2

    work page 2024

Show all 81 references
  1. [9]

    Gpg: A simple and strong reinforce- ment learning baseline for model reasoning

    Xiangxiang Chu, Hailang Huang, Xiao Zhang, Fei Wei, and Yong Wang. Gpg: A simple and strong reinforce- ment learning baseline for model reasoning. arXiv preprint arXiv:2504.02546, 2025. 1

    2025

  2. [10]

    The cityscapes dataset for semantic urban scene understanding

    Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern reco...

    2016

  3. [11]

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

    Muhammad Sohail Danish, Muhammad Haris Khan, Muhammad Akhtar Munir, M Saquib Sarfraz, and Mohsen Ali. Improving single domain-generalized object detection: A focus on diversification and alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogn...

    2024

  4. [12]

    Learning to prompt for open-vocabulary object detection with vision-language model

    Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14084–14093, 2022. 2, 5, 6, 1

    2022

  5. [13]

    Boosting object detection with zero-shot day-night domain adaptation

    Zhipeng Du, Miaojing Shi, and Jiankang Deng. Boosting object detection with zero-shot day-night domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12666–12676, 2024. 2, 6, 7, 8

    2024

  6. [14]

    Poda: Prompt-driven zero- shot domain adaptation

    Mohammad Fahes, Tuan-Hung Vu, Andrei Bursuc, Patrick P´erez, and Raoul De Charette. Poda: Prompt-driven zero- shot domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18623– 18633, 2023. 1, 2, 4, 6, 7, 8

    2023

  7. [15]

    Rapid: Zero-shot domain adaptation for code search with pre-trained models

    Guodong Fan, Shizhan Chen, Cuiyun Gao, Jianmao Xiao, Tao Zhang, and Zhiyong Feng. Rapid: Zero-shot domain adaptation for code search with pre-trained models. ACM Transactions on Software Engineering and Methodology , 33(5):1–35, 2024. 2

    2024

  8. [16]

    Domain adaptation via prompt learning

    Chunjiang Ge, Rui Huang, Mixue Xie, Zihang Lai, Shiji Song, Shuang Li, and Gao Huang. Domain adaptation via prompt learning. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2

    2023

  9. [17]

    Are we ready for autonomous driving? the kitti vision benchmark suite

    Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354–3361. IEEE, 2012. 6

    2012

  10. [18]

    Generative adversarial nets

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2

    2014

  11. [19]

    Vehicle detection and tracking in ad- verse weather using a deep learning framework

    Mahmoud Hassaballah, Mourad A Kenk, Khan Muhammad, and Shervin Minaee. Vehicle detection and tracking in ad- verse weather using a deep learning framework. IEEE trans- actions on intelligent transportation systems , 22(7):4230– 4242, 2020. 6

    2020

  12. [20]

    Generalized diffusion detector: Mining robust features from diffusion models for domain-generalized de- tection

    Boyong He, Yuxiang Ji, Qianwen Ye, Zhuoyue Tan, and Liaoni Wu. Generalized diffusion detector: Mining robust features from diffusion models for domain-generalized de- tection. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 9921–9932, 2025. 3

    2025

  13. [21]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed- ings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 6

    2016

  14. [22]

    Unsu- pervised prompt tuning for text-driven object detection

    Weizhen He, Weijie Chen, Binbin Chen, Shicai Yang, Di Xie, Luojun Lin, Donglian Qi, and Yueting Zhuang. Unsu- pervised prompt tuning for text-driven object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2651–2661, 2023. 2

    2023

  15. [23]

    Progressive domain adaptation for object detection

    Han-Kai Hsu, Chun-Han Yao, Yi-Hsuan Tsai, Wei-Chih Hung, Hung-Yu Tseng, Maneesh Singh, and Ming-Hsuan Yang. Progressive domain adaptation for object detection. In Proceedings of the IEEE/CVF winter conference on applica- tions of computer vision, pages 749–757, 2020. 1

    2020

  16. [24]

    Arbitrary style transfer in real-time with adaptive instance normalization

    Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceed- ings of the IEEE international conference on computer vision, pages 1501–1510, 2017. 4

    2017

  17. [25]

    Diffusion features to bridge domain gap for semantic segmentation

    Yuxiang Ji, Boyong He, Chenyuan Qu, Zhuoyue Tan, Chuan Qin, and Liaoni Wu. Diffusion features to bridge domain gap for semantic segmentation. In ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE, 2025. 3

    2025

  18. [26]

    Driving in the matrix: Can virtual worlds replace human- generated annotations for real world tasks? arXiv preprint arXiv:1610.01983, 2016

    Matthew Johnson-Roberson, Charles Barto, Rounak Mehta, Sharath Nittur Sridhar, Karl Rosaen, and Ram Vasudevan. Driving in the matrix: Can virtual worlds replace human- generated annotations for real world tasks? arXiv preprint arXiv:1610.01983, 2016. 6

    arXiv

  19. [27]

    Vilt: Vision- and-language transformer without convolution or region su- pervision

    Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision- and-language transformer without convolution or region su- pervision. In International conference on machine learning, pages 5583–5594. PMLR, 2021. 1

    2021

  20. [28]

    Zero- shot deep domain adaptation with common representation learning

    Mohammed Kutbi, Kuan-Chuan Peng, and Ziyan Wu. Zero- shot deep domain adaptation with common representation learning. IEEE Transactions on Pattern Analysis and Ma- chine Intelligence, 44(7):3909–3924, 2021. 2

    2021

  21. [29]

    Object-aware domain generalization for object detection

    Wooju Lee, Dasol Hong, Hyungtae Lim, and Hyun Myung. Object-aware domain generalization for object detection. In Proceedings of the AAAI Conference on Artificial Intelli- gence, volume 38, pages 2947–2955, 2024. 2, 6, 7, 8, 4

    2024

  22. [30]

    Zero-shot day-night domain adaptation with a physics prior

    Attila Lengyel, Sourav Garg, Michael Milford, and Jan C van Gemert. Zero-shot day-night domain adaptation with a physics prior. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4399–4409, 2021. 2

    2021

  23. [31]

    Prompt-driven dynamic object-centric learning for single do- main generalization

    Deng Li, Aming Wu, Yaowei Wang, and Yahong Han. Prompt-driven dynamic object-centric learning for single do- main generalization. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 17606–17615, 2024. 6, 7, 2, 3

    2024

  24. [32]

    Learning domain-aware detection head with prompt tuning

    Haochen Li, Rui Zhang, Hantao Yao, Xinkai Song, Yifan Hao, Yongwei Zhao, Ling Li, and Yunji Chen. Learning domain-aware detection head with prompt tuning. Advances in Neural Information Processing Systems, 36, 2024. 2, 4

    2024

  25. [33]

    Da-ada: Learning domain-aware adapter for domain adaptive object detection

    Haochen Li, Rui Zhang, Hantao Yao, Xin Zhang, Yifan Hao, Xinkai Song, Xiaqing Li, Yongwei Zhao, Yunji Chen, and Ling Li. Da-ada: Learning domain-aware adapter for domain adaptive object detection. Advances in Neural Information Processing Systems, 37:103574–103598, 2024. 1

    2024

  26. [34]

    Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Interna- tional conference on machine learning, pages 12888–12900. PMLR, 2022. 1

    2022

  27. [35]

    Align be- fore fuse: Vision and language representation learning with momentum distillation

    Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align be- fore fuse: Vision and language representation learning with momentum distillation. Advances in neural information pro- cessing systems, 34:9694–9705, 2021. 1

    2021

  28. [36]

    A comprehensive survey on source-free domain adap- tation

    Jingjing Li, Zhiqi Yu, Zhekai Du, Lei Zhu, and Heng Tao Shen. A comprehensive survey on source-free domain adap- tation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1

    2024

  29. [37]

    Grounded language- image pre-training

    Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jian- wei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language- image pre-training. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , ...

    2022

  30. [38]

    Next token is enough: Realistic image quality and aesthetic scoring with multimodal large language model

    Mingxing Li, Rui Wang, Lei Sun, Yancheng Bai, and Xi- angxiang Chu. Next token is enough: Realistic image quality and aesthetic scoring with multimodal large language model. arXiv preprint arXiv:2503.06141, 2025. 1

    2025 arXiv

  31. [39]

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

    Yajing Liu, Shijun Zhou, Xiyao Liu, Chunhui Hao, Baojie Fan, and Jiandong 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, pages 28838–28847, 2024. 2, 6, 7

    2024

  32. [40]

    Similarity min-max: Zero-shot day-night domain adap- tation

    Rundong Luo, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Similarity min-max: Zero-shot day-night domain adap- tation. In Proceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 8104–8114, 2023. 2

    2023

  33. [41]

    Confmix: Unsupervised domain adaptation for object detection via confidence-based mixing

    Giulio Mattolin, Luca Zanella, Elisa Ricci, and Yiming Wang. Confmix: Unsupervised domain adaptation for object detection via confidence-based mixing. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 423–433, 2023. 1

    2023

  34. [42]

    Domain-aware vi- sual bias eliminating for generalized zero-shot learning

    Shaobo Min, Hantao Yao, Hongtao Xie, Chaoqun Wang, Zheng-Jun Zha, and Yongdong Zhang. Domain-aware vi- sual bias eliminating for generalized zero-shot learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12664–12673, 2020. 1, 2

    2020

  35. [43]

    Learning cnn on vit: A hybrid model to explicitly class-specific boundaries for domain adaptation

    Ba Hung Ngo, Nhat-Tuong Do-Tran, Tuan-Ngoc Nguyen, Hae-Gon Jeon, and Tae Jong Choi. Learning cnn on vit: A hybrid model to explicitly class-specific boundaries for domain adaptation. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages ...

    2024

  36. [44]

    Unsupervised domain adap- tation of object detectors: A survey

    Poojan Oza, Vishwanath A Sindagi, Vibashan Vishnukumar Sharmini, and Vishal M Patel. Unsupervised domain adap- tation of object detectors: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1

    2023

  37. [45]

    Zero-shot deep domain adaptation

    Kuan-Chuan Peng, Ziyan Wu, and Jan Ernst. Zero-shot deep domain adaptation. In Proceedings of the European Con- ference on Computer Vision (ECCV), pages 764–781, 2018. 2

    2018

  38. [46]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. In International conference on machine learning, ...

    2021

  39. [47]

    Srcd: Semantic reasoning with compound domains for single-domain generalized object de- tection

    Zhijie Rao, Jingcai Guo, Luyao Tang, Yue Huang, Xing- hao Ding, and Song Guo. Srcd: Semantic reasoning with compound domains for single-domain generalized object de- tection. arXiv preprint arXiv:2307.01750, 2023. 2

    2023 arXiv

  40. [48]

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

    Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information process- ing systems, 28, 2015. 5, 6, 7, 8

    2015

  41. [49]

    Se- mantic foggy scene understanding with synthetic data

    Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Se- mantic foggy scene understanding with synthetic data. In- ternational Journal of Computer Vision, 126:973–992, 2018. 6

    2018

  42. [50]

    Ad-clip: Adapting domains in prompt space using clip

    Mainak Singha, Harsh Pal, Ankit Jha, and Biplab Banerjee. Ad-clip: Adapting domains in prompt space using clip. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4355–4364, 2023. 2, 3

    2023

  43. [51]

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

    Vidit Vidit, Martin Engilberge, and Mathieu Salzmann. Clip the gap: A single domain generalization approach for object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3219–3229,

  44. [52]

    Instance relation graph guided source-free domain adaptive object de- tection

    Vibashan VS, Poojan Oza, and Vishal M Patel. Instance relation graph guided source-free domain adaptive object de- tection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3520–3530,

  45. [53]

    Pond: Multi-source time se- ries domain adaptation with information-aware prompt tun- ing

    Junxiang Wang, Guangji Bai, Wei Cheng, Zhengzhang Chen, Liang Zhao, and Haifeng Chen. Pond: Multi-source time se- ries domain adaptation with information-aware prompt tun- ing. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages 3140–3151,

  46. [54]

    Do- main shift preservation for zero-shot domain adaptation

    Jinghua Wang, Ming-Ming Cheng, and Jianmin Jiang. Do- main shift preservation for zero-shot domain adaptation. IEEE Transactions on Image Processing , 30:5505–5517,

  47. [55]

    Conditional coupled gen- erative adversarial networks for zero-shot domain adaptation

    Jinghua Wang and Jianmin Jiang. Conditional coupled gen- erative adversarial networks for zero-shot domain adaptation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3375–3384, 2019. 2

    2019

  48. [56]

    Adversarial learning for zero-shot domain adaptation

    Jinghua Wang and Jianmin Jiang. Adversarial learning for zero-shot domain adaptation. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXI 16 , pages 329–344. Springer, 2020

    2020

  49. [57]

    Learning across tasks for zero-shot domain adaptation from a single source domain

    Jinghua Wang and Jianmin Jiang. Learning across tasks for zero-shot domain adaptation from a single source domain. IEEE Transactions on Pattern Analysis and Machine Intelli- gence, 44(10):6264–6279, 2021. 2

    2021

  50. [58]

    Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework

    Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. InInternational conference on machine learning,...

    2022

  51. [59]

    Simvlm: Simple visual language model pretraining with weak supervision

    Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. Simvlm: Simple visual language model pretraining with weak supervision. In International Conference on Learning Representations, 2022. 4

    2022

  52. [60]

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

    Aming Wu and Cheng 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, pages 847–856,

  53. [61]

    G-nas: Generalizable neu- ral architecture search for single domain generalization object detection

    Fan Wu, Jinling Gao, Lanqing Hong, Xinbing Wang, Chenghu Zhou, and Nanyang Ye. G-nas: Generalizable neu- ral architecture search for single domain generalization object detection. arXiv preprint arXiv:2402.04672, 2024. 2, 7

    2024 arXiv

  54. [62]

    Hgnet: Hybrid generative network for zero-shot domain adaptation

    Haifeng Xia and Zhengming Ding. Hgnet: Hybrid generative network for zero-shot domain adaptation. In European Con- ference on Computer Vision , pages 55–70. Springer, 2020. 1

    2020

  55. [63]

    Mmap: Multi-modal alignment prompt for cross- domain multi-task learning

    Yi Xin, Junlong Du, Qiang Wang, Ke Yan, and Shouhong Ding. Mmap: Multi-modal alignment prompt for cross- domain multi-task learning. InProceedings of the AAAI Con- ference on Artificial Intelligence, volume 38, pages 16076– 16084, 2024. 2

    2024

  56. [64]

    Hs-star: Hierarchical sampling for self-taught reasoners via difficulty estimation and budget reallocation, 2025

    Feng Xiong, Hongling Xu, Yifei Wang, Runxi Cheng, Yong Wang, and Xiangxiang Chu. Hs-star: Hierarchical sampling for self-taught reasoners via difficulty estimation and budget reallocation, 2025. 2

    2025

  57. [65]

    Unified language-driven zero-shot domain adaptation

    Senqiao Yang, Zhuotao Tian, Li Jiang, and Jiaya Jia. Unified language-driven zero-shot domain adaptation. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23407–23415, 2024. 1, 2

    2024

  58. [66]

    Bdd100k: A diverse driving dataset for heterogeneous multi- task learning

    Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multi- task learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p...

  59. [67]

    Coca: Contrastive captioners are image-text foundation models

    Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mo- jtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 4

    2022 arXiv

  60. [68]

    Learning domain invariant prompt for vision-language models

    Cairong Zhao, Yubin Wang, Xinyang Jiang, Yifei Shen, Kaitao Song, Dongsheng Li, and Duoqian Miao. Learning domain invariant prompt for vision-language models. IEEE Transactions on Image Processing, 2024. 2, 3, 4

    2024

  61. [69]

    Scene-adaptive and region-aware multi-modal prompt for open vocabulary object detection

    Xiaowei Zhao, Xianglong Liu, Duorui Wang, Yajun Gao, and Zhide Liu. Scene-adaptive and region-aware multi-modal prompt for open vocabulary object detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16741–16750, 2024. 2 Appendix...

    2024

  62. [71]

    - ✓ ✓ ✓ 37.8 38.3 17.1 32.2

  63. [72]

    ✓ - ✓ ✓ 38.5 38.7 16.9 32.8

  64. [73]

    ✓ ✓ - ✓ 39.2 39.8 18.5 33.1

  65. [74]

    ✓ ✓ ✓ - 38.9 39.5 18.3 32.8

  66. [75]

    ✓ ✓ - - 35.1 36.2 16.7 30.5

  67. [76]

    - ✓ ✓ - 35.6 36.1 16.3 29.9

  68. [77]

    - ✓ - ✓ 35.9 36.5 16.8 30.5

  69. [78]

    ✓ - - ✓ 32.9 34.8 14.2 27.5

  70. [79]

    ✓ - ✓ - 32.8 34.3 14.1 27.2

  71. [80]

    ✓ - - - 32.1 34.0 13.5 26.5

  72. [81]

    ✓ ✓ ✓ ✓ 40.0 41.5 19.8 34.5 Table 10. Ablation study of internal modules.𝑃𝑟𝑜𝑚𝑝𝑡 denotes the proposed domain adaptation prompt,𝐸𝑛ℎ𝑎𝑛𝑐𝑒 represents the uni- fied representation enhancement,𝐼𝑚𝑔 indicates the relative domain distance strategy, and𝐼𝑛𝑠 signifies the positive-negative...

  73. [2023]

    2, 4, 5, 6, 7, 8, 1, 3

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.