Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

Diffusion Domain Teacher: Diffusion Guided Domain Adaptive Object Detector

T0 review · 3 major / 6 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 A frozen Stable Diffusion U-Net, used as a detector backbone and then as a pseudo-label teacher, lifts cross-domain detection mAP by 21.2% over baseline and beats prior methods by 5.7% across six benchmarks.

desk verdict Frozen diffusion backbone as a pseudo-label teacher is a genuinely new idea with strong empirical support, but the SOTA margin is inflated by test-set-tuned hyperparameters. read the letter →

arxiv 2506.04211 v1 pith:RR5CUJB7 submitted 2025-06-04 cs.CV

Boyong He , Yuxiang Ji , Zhuoyue Tan , Liaoni Wu This is my paper · ORCID
classification cs.CV
keywords unsuperviseddomainadaptationcross-domainobjectdetectiondiffusionmodelstablepseudo-labelself-trainingmeanteacherfrozenbackboneFasterR-CNN
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 a diffusion model—normally thought of as an image generator—can do something an object-detection backbone is not expected to do: transfer across visual domains. It builds a detector whose backbone is a frozen Stable Diffusion U-Net, trained only on labeled source images, and then uses that detector as a teacher that writes pseudo-labels on unlabeled target images. A student detector is trained on those pseudo-labels while a mean teacher, updated by exponential moving average, produces the final model. The reported result is an average 21.2% mAP improvement over the source-only baseline across six datasets, and an average 5.7% mAP improvement over prior methods. Because the diffusion weights stay frozen and are not used at deployment, the method changes training but not inference cost.

What carries the argument

The load-bearing mechanism is a frozen diffusion U-Net used as a hierarchical feature extractor. For each image, noise is added to move it to a time step t; the U-Net's upsampling activations at four stages, concatenated across several saved time steps, are passed through a small trainable bottleneck that projects them to [256, 512, 1024, 2048] channels, matching ResNet's feature pyramid so that Faster R-CNN can consume them. The same diffusion detector then acts as a pseudo-label teacher in a mean-teacher self-training loop: weakly augmented target images are labeled, a threshold sigma keeps confident boxes, and the student is trained on those pseudo-labels with strong augmentation plus the supervised source loss; EMA updates the mean teacher, which is used for final inference.

What would settle it

Run DDT on a fresh target domain with sigma=0.5, lambda=1, and five time steps fixed before seeing any target labels, and compare against the same self-training loop with a fully trainable ResNet-101 teacher. If the gap to the ResNet teacher nearly vanishes, the 'frozen diffusion is a better teacher' claim fails; if the gap persists, the claim survives.

Watch

Extended reading notes

Core claim

The central discovery, stated on the paper's own terms, is that a frozen diffusion U-Net is a better cross-domain teacher than fully trainable backbones. The paper supports this claim by comparing the diffusion teacher against ResNet, ConvNeXt, Swin, ViT, MAE, and GLIP as alternative teachers: although the diffusion detector underperforms several backbones on intra-domain tasks, it gives the largest gains to a ResNet-101 student on three adaptations, and it consistently improves students built on the other backbones. The authors conclude that supervised training on the source domain makes ordinary backbones homogeneous in their optimization, while the frozen diffusion representation provides more diverse and generalizable guidance to the student.

Load-bearing premise

The headline gains assume the three settings used to report them—pseudo-label threshold 0.5, unsupervised-loss weight 1, and five diffusion time steps with five saved steps—still work when chosen without peeking at the target test set, since the paper selects them by ablation on those same test sets.

Editorial extensions

If this is right

  • If correct, any deployment-stage detector can be kept unchanged: the diffusion teacher is used only during training, so the final model runs at the student's normal speed.
  • If correct, even very strong pretrained backbones (ConvNeXt, Swin, ViT, MAE, GLIP) gain from diffusion-provided pseudo-labels, implying these features are complementary to supervised and language-supervised representations.
  • If correct, the method is most valuable where the domain gap is largest: the paper reports relative improvements of 95% on Clipart, 163% on Comic, and 48% on Watercolor over baseline.
  • If correct, both parts of the framework matter: ablations show removing the mean teacher costs 3.1 mAP on average and removing the diffusion teacher costs 6.7 mAP, so neither component is decorative.

Reading between the lines

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

  • Beyond the paper: because the diffusion teacher's advantage grows with domain gap in the reported real-to-artistic results, a testable prediction is that gains are monotone in target style divergence; an interpolation study along a style axis would settle it.
  • Beyond the paper: the frozen U-Net is never exposed to target data, so its pseudo-labels come from a fixed representation; adapting only the bottleneck or detection head, or adding a feature-space consistency loss, is a natural next step the paper does not explore.
  • Beyond the paper: the selected pseudo-label threshold and loss weight were tuned on the same target test sets used for reporting, so the 21.2% figure may shrink under a protocol that fixes these knobs before seeing the target domain.
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 / 6 minor

Summary. The paper proposes Diffusion Domain Teacher (DDT), an unsupervised domain adaptation method for object detection. The authors freeze a Stable Diffusion U-Net, extract multi-scale features from its upsampling path through a trainable bottleneck, and train a Faster R-CNN detector on the labeled source domain. This 'diffusion detector' is then used as a teacher to generate pseudo-labels on the unlabeled target domain, while a student detector with a conventional backbone is trained on the pseudo-labels together with source data in a mean-teacher self-training loop (EMA teacher). Experiments cover six adaptation tasks across three benchmarks (Cityscapes/BDD100K, Sim10K, VOC/Clipart/Comic/Watercolor). The authors report an average 21.2 mAP improvement over a source-only baseline and claim an average 5.7 mAP advantage over previous SOTA. Additional ablations examine the choice of teacher (compared with ConvNeXt, Swin, ViT, MAE, GLIP), the effect of diffusion time steps, pseudo-label threshold, loss weight, augmentation, and Stable Diffusion versions, and the appendix provides FCOS results.

Significance. The core idea is clear and the paper is generally well organized. The strongest contribution is the empirical finding that a frozen diffusion-based detector, despite being a relatively weak intra-domain feature extractor (Table 7), serves as a more effective pseudo-label teacher than fully trainable strong backbones (Table 8), and that the student benefits substantially from the resulting pseudo-labels. The evaluation covers six datasets and compares with many recent methods, and the authors release code. The main limitations are the selection of critical hyperparameters on the target test sets (Tables 10-12), the absence of multiple seeds, and an inconsistent SOTA comparison; these issues make the exact 5.7 mAP headline margin uncertain, although the overall direction of the improvement over the source-only baseline appears robust.

major comments (3)
  1. [Section 4.5, Appendix A.1, Tables 10-12] The default values of the diffusion time steps/save steps, pseudo-label threshold sigma, and unsupervised loss weight lambda are selected by ablating on the same target test sets used in the final evaluation, without a held-out target validation split. For example, Table 12 shows that on V->Ca, sigma=0.6 yields 56.1 mAP versus 55.6 for the chosen sigma=0.5, and on S->Cs, sigma=0.4 yields 64.2 versus 64.0; Table 10 shows that on V->Co and V->W, the 20/10 step schedule gives 41.3 and 54.6 versus 39.4 and 53.8 for the default 5/5 schedule. Since the reported SOTA margin of 5.7 mAP depends on these choices, the headline comparison is a selection result rather than an unbiased estimate. The authors should either fix the hyperparameters using a target validation split (or source-only validation), or report sensitivity across a range of settings and adjust the claim accordingly.
  2. [Section 3.2] The diffusion feature extractor is described only at a high level: it is not stated which Stable Diffusion version (SD-1.5 or SD-2.1) is used for the main results, how the features from the four upsampling stages are aggregated across time steps, or what the bottle-neck structure contains (number and type of layers, output strides, normalization). These details are necessary to reproduce the method, since the entire contribution rests on this feature extractor. Please provide a full specification (or a clear pointer to the released code with the relevant configuration).
  3. [Section 4.4, Tables 1-6] All quantitative results are reported as single-run numbers, and the comparison to prior work is not controlled for detector/backbone: e.g., in Table 3 DDT with FRCNN-R101 reaches 64.0 mAP while HT uses FCOS-V16 and reaches 65.5, and in Table 1 DDT is compared against HT with FCOS-V16. Without multiple seeds and without a matched-backbone comparison, the claimed average 5.7 mAP advantage over SOTA is not statistically grounded. Please report mean and standard deviation over at least three seeds for the main tables, and, where possible, include the same backbone/detector as the strongest prior method.
minor comments (6)
  1. [Abstract] There is a typo in 'Real2Artistic}' (unbalanced brace).
  2. [Throughout] There are several typos, e.g., 'Artifcial', 'domian', 'Fozen-Diffusion', and 'teak' in Section 2.1; please proofread the manuscript.
  3. [Table 14] The text says SD-2.1 is the latest version; please specify the release date and clarify why SD-1.5 is preferred, especially since Table 14 shows SD-2.1 is better on S->B.
  4. [Figure 1] The caption says 'Evaluation results on Clipart' but the figure appears to show a training curve that may be for a different target dataset; please verify the caption matches the plotted experiment.
  5. [Section 3.2] The term 'inversion process' is used for adding noise to the input image; this is actually the forward diffusion process. Please align the terminology with the diffusion model literature.
  6. [Tables 7 and 8] The 'Rel.(%)' column is not defined; please state the formula used (e.g., cross-domain mAP divided by intra-domain mAP) so the reader can interpret the relative cross-domain capability.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the DDT results are empirical comparisons on external benchmarks, not derivations from their own inputs.

full rationale

The central claim is an empirical one: that a frozen Stable Diffusion U-Net, used as a detection backbone and then as a pseudo-label teacher in a mean-teacher self-training loop, improves cross-domain detection. The evidence consists of mAP measurements on six standard target test sets, compared against external published methods (HT, AT, CMT, etc.) and independently pretrained backbones (ResNet, ConvNeXt, Swin, ViT, MAE, GLIP). No equation in the paper defines the reported mAP in terms of its own ingredients: the loss in Eq. (5), L = L_sup + λ·L_unsup, is a standard supervised-plus-pseudo-label objective, and the pseudo-labels come from a detector trained on the source domain, not from the evaluation labels. The frozen diffusion model is an externally pretrained artifact (SD-1.5/SD-2.1), and the claimed gains are measured outcomes rather than predictions forced by construction. The paper does select hyperparameters (σ, λ, time steps) by ablating on the same target test sets used for final reporting (Tables 10-12), which is a benchmark-selection weakness and a legitimate correctness risk, but it is not circularity in the derivation-chain sense: the method's improvement over its own baseline and over SOTA is still an empirical result. There are no load-bearing self-citations, no uniqueness theorem imported from the authors, and no fitted parameter renamed as a prediction. Accordingly, the paper is self-contained against external benchmarks and receives a circularity score of 0.

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

The central claim rests on external pretrained diffusion features, standard self-training assumptions, and a set of hyperparameters selected by ablating on the target test sets. No new physical entity or fitted constant is introduced; the paper is an empirical engineering contribution.

free parameters (4)
  • Pseudo-label confidence threshold sigma = 0.5
    Chosen by ablation on Cs->B, S->Cs, and V->Ca (Table 12) rather than a held-out validation set.
  • Unsupervised loss weight lambda = 1.0
    Tuned via ablation on three target test sets (Table 11), then fixed.
  • Diffusion time steps and save steps = 5 and 5
    Selected from Table 10 as an accuracy/efficiency trade-off after observing target-test results.
  • EMA decay alpha = 0.999
    Taken from prior Mean Teacher practice, not independently validated here.
assumptions (4)
  • domain assumption Unlabeled target images share the same object categories as the labeled source set and annotations align across datasets.
    The method assumes pseudo-label classes from source transfer to target; this holds by benchmark construction but is an assumption about real UDA.
  • domain assumption Pretrained Stable Diffusion (SD-1.5) U-Net features are available and frozen, and its training distribution overlaps enough with source and target detection data.
    The whole method relies on an externally pretrained generative model, not trained for detection; Section 3.2.
  • standard math The mean-teacher EMA update stabilizes pseudo-label training, as originally justified in Mean Teacher [71].
    Adopted as prior methodology, not derived; Section 3.3, Eq. 4.
  • domain assumption Standard stochastic optimization and augmentation settings suffice to reproduce the reported mAP.
    Implementation details in Section 4.3 rely on MMDetection defaults; no seed variance is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Diffusion Domain Teacher: Diffusion Guided Domain Adaptive Object Detector." pith.science (2026). https://pith.science/paper/RR5CUJB7

@misc{pith2026250604211,
  author       = {Pith},
  title        = {Pith review of: Diffusion Domain Teacher: Diffusion Guided Domain Adaptive Object Detector},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RR5CUJB7}},
  note         = {Machine review of arXiv:2506.04211}
}
read the original abstract

Object detectors often suffer a decrease in performance due to the large domain gap between the training data (source domain) and real-world data (target domain). Diffusion-based generative models have shown remarkable abilities in generating high-quality and diverse images, suggesting their potential for extracting valuable feature from various domains. To effectively leverage the cross-domain feature representation of diffusion models, in this paper, we train a detector with frozen-weight diffusion model on the source domain, then employ it as a teacher model to generate pseudo labels on the unlabeled target domain, which are used to guide the supervised learning of the student model on the target domain. We refer to this approach as Diffusion Domain Teacher (DDT). By employing this straightforward yet potent framework, we significantly improve cross-domain object detection performance without compromising the inference speed. Our method achieves an average mAP improvement of 21.2% compared to the baseline on 6 datasets from three common cross-domain detection benchmarks (Cross-Camera, Syn2Real, Real2Artistic}, surpassing the current state-of-the-art (SOTA) methods by an average of 5.7% mAP. Furthermore, extensive experiments demonstrate that our method consistently brings improvements even in more powerful and complex models, highlighting broadly applicable and effective domain adaptation capability of our DDT. The code is available at https://github.com/heboyong/Diffusion-Domain-Teacher.

Figures

Figures reproduced from arXiv: 2506.04211 by the authors.

Figure 1
Figure 1. Evaluation results on Clipart [30] during training. It is evident that the performance of the student significantly im￾proves after entering self-training, even surpassing the diffusion teacher, and the mean teacher exhibits better performance com￾pared to the student. has been some effort to adapt diffusion models for image generation and manipulation, as seen with tools like LoRA [29] and Control￾Net [83], there i… view at source ↗
Figure 2
Figure 2. Overview of our proposed Diffusion Domain Teacher (DDT). Left: [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Qualitative prediction results and feature visualization of baseline and our DDT. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Error analysis on Clipart. It is evident that our method significantly reduces false negatives, which correspond to missed detections. of Resnet50 and ResNet101 significantly surpass the previous best method AT [44] by 3.9 and 4.8 mAP, respectively. On Comic [30], the …
Figure 5
Figure 5. Figure 5: Main content of our work. Left: We present three cross-domain detection benchmarks and visualize the image [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: Qualitative prediction results and feature visualization of baseline and our DDT from Cityscapes to BDD100K. [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]
Figure 7
Figure 7. Figure 7: Qualitative prediction results and feature visualization of baseline and our DDT from Sim10K to BDD100K. [PITH_FULL_IMAGE:figures/full_fig_p019_7.png]
Figure 8
Figure 8. Figure 8: Qualitative prediction results and feature visualization of baseline and our DDT from Sim10K to Cityscapes. [PITH_FULL_IMAGE:figures/full_fig_p019_8.png]
Figure 9
Figure 9. Figure 9: Qualitative prediction results and feature visualization of baseline and our DDT from VOC to Clipart. [PITH_FULL_IMAGE:figures/full_fig_p020_9.png]
Figure 10
Figure 10. Figure 10: Qualitative prediction results and feature visualization of baseline and our DDT from VOC to Comic. [PITH_FULL_IMAGE:figures/full_fig_p021_10.png]
Figure 11
Figure 11. Figure 11: Qualitative prediction results and feature visualization of baseline and our DDT from VOC to Watercolor. [PITH_FULL_IMAGE:figures/full_fig_p022_11.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. Conditional Diffusion Guided Knowledge Transfer for Multi-Domain Knowledge Graph Completion

    cs.CL 2026-07 accept novelty 7.0 of 10

    Conditional diffusion generates unbiased domain-general entity embeddings from support KGs, lifting multi-domain KG completion by 4.3% average MRR over prior consistency methods.

Reference graph

Works this paper leans on

96 extracted references · 52 canonical work pages · cited by 1 Pith paper

  1. [1]

    Dmitry Baranchuk, Andrey Voynov, Ivan Rubachev, Valentin Khrulkov, and Artem Babenko. 2022. Label-Efficient Semantic Segmentation with Diffusion Models. In International Conference on Learning Representations

    2022

  2. [2]

    Zhaowei Cai and Nuno Vasconcelos. 2018. Cascade r-cnn: Delving into high quality object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition. 6154–6162

    2018

  3. [3]

    Shengcao Cao, Dhiraj Joshi, Liang-Yan Gui, and Yu-Xiong Wang. 2023. Con- trastive mean teacher for domain adaptive object detectors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 23839–23848

    2023

  4. [4]

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexan- der Kirillov, and Sergey Zagoruyko. 2020. End-to-end object detection with transformers. In European conference on computer vision . Springer, 213–229

    2020

  5. [5]

    Chaoqi Chen, Jiongcheng Li, Zebiao Zheng, Yue Huang, Xinghao Ding, and Yizhou Yu. 2021. Dual bipartite graph learning: A general approach for domain adaptive object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision. 2703–2712

    2021

  6. [7]

    Chaoqi Chen, Zebiao Zheng, Xinghao Ding, Yue Huang, and Qi Dou. 2020. Har- monizing transferability and discriminability for adapting object detectors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 8869–8878

    2020

  7. [8]

    Chaoqi Chen, Zebiao Zheng, Yue Huang, Xinghao Ding, and Yizhou Yu. 2021. I3net: Implicit instance-invariant network for adapting one-stage object detec- tors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 12576–12585

    2021

  8. [9]

    Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. 2019. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155 (2019)

    arXiv 2019

Show all 96 references
  1. [10]

    Meilin Chen, Weijie Chen, Shicai Yang, Jie Song, Xinchao Wang, Lei Zhang, Yunfeng Yan, Donglian Qi, Yueting Zhuang, Di Xie, et al. 2022. Learning Domain Adaptive Object Detection with Probabilistic Teacher. InInternational Conference on Machine Learning. PMLR, 3040–3055

    2022

  2. [11]

    Yuhua Chen, Wen Li, Christos Sakaridis, Dengxin Dai, and Luc Van Gool. 2018. Domain adaptive faster r-cnn for object detection in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition . 3339–3348

    2018

  3. [12]

    Yuhua Chen, Haoran Wang, Wen Li, Christos Sakaridis, Dengxin Dai, and Luc Van Gool. 2021. Scale-aware domain adaptive faster r-cnn. International Journal of Computer Vision 129, 7 (2021), 2223–2243

    2021

  4. [13]

    Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus En- zweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. 2016. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and patt...

    2016

  5. [14]

    Jinhong Deng, Wen Li, Yuhua Chen, and Lixin Duan. 2021. Unbiased mean teacher for cross-domain object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 4091–4101

    2021

  6. [15]

    Jinhong Deng, Dongli Xu, Wen Li, and Lixin Duan. 2023. Harmonious teacher for cross-domain object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 23829–23838

    2023

  7. [16]

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xi- aohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. 2020. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv prepri...

    2020 arXiv

  8. [17]

    Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. 2010. The pascal visual object classes (voc) challenge. Inter- national journal of computer vision 88 (2010), 303–338

    2010

  9. [18]

    Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario March, and Victor Lempitsky. 2016. Domain-adversarial training of neural networks. Journal of machine learning research 17, 59 (2016), 1–35

    2016

  10. [19]

    Kaixiong Gong, Shuang Li, Shugang Li, Rui Zhang, Chi Harold Liu, and Qiang Chen. 2022. Improving transferability for domain adaptive detection transformers. In Proceedings of the 30th ACM International Conference on Multimedia . 1543– 1551

    2022

  11. [20]

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2020. Generative adversarial networks. Commun. ACM 63, 11 (2020), 139–144

    2020

  12. [21]

    Dayan Guan, Jiaxing Huang, Aoran Xiao, Shijian Lu, and Yanpeng Cao. 2021. Uncertainty-aware unsupervised domain adaptation in object detection. IEEE Transactions on Multimedia 24 (2021), 2502–2514

    2021

  13. [22]

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick

  14. [23]

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

    2016

  15. [24]

    Mengzhe He, Yali Wang, Jiaxi Wu, Yiru Wang, Hanqing Li, Bo Li, Weihao Gan, Wei Wu, and Yu Qiao. 2022. Cross domain object detection by target-perceived dual branch distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 9570–9580

    2022

  16. [25]

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising diffusion probabilistic models. Advances in neural information processing systems 33 (2020), 6840–6851

    2020

  17. [26]

    Lukas Hoyer, Dengxin Dai, Haoran Wang, and Luc Van Gool. 2023. MIC: Masked image consistency for context-enhanced domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 11721–11732

    2023

  18. [27]

    Cheng-Chun Hsu, Yi-Hsuan Tsai, Yen-Yu Lin, and Ming-Hsuan Yang. 2020. Ev- ery pixel matters: Center-aware feature alignment for domain adaptive object detector. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part IX 16 . S...

    2020

  19. [28]

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

    2020

  20. [29]

    Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. 2022. LoRA: Low-Rank Adaptation of Large Language Models. In International Conference on Learning Representations

    2022

  21. [30]

    Naoto Inoue, Ryosuke Furuta, Toshihiko Yamasaki, and Kiyoharu Aizawa. 2018. Cross-domain weakly-supervised object detection through progressive domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition. 5001–5009

    2018

  22. [31]

    Junguang Jiang, Baixu Chen, Jianmin Wang, and Mingsheng Long. 2021. Decou- pled Adaptation for Cross-Domain Object Detection. In International Conference on Learning Representations

    2021

  23. [32]

    Matthew Johnson-Roberson, Charles Barto, Rounak Mehta, Sharath Nittur Srid- har, Karl Rosaen, and Ram Vasudevan. 2017. Driving in the Matrix: Can virtual worlds replace human-generated annotations for real world tasks?. In 2017 IEEE International Conference on Robotics and Aut...

    2017

  24. [33]

    Seunghyeon Kim, Jaehoon Choi, Taekyung Kim, and Changick Kim. 2019. Self- training and adversarial background regularization for unsupervised domain adaptive one-stage object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 6092–6101

    2019

  25. [34]

    Taekyung Kim, Minki Jeong, Seunghyeon Kim, Seokeon Choi, and Changick Kim

  26. [35]

    Congcong Li, Dawei Du, Libo Zhang, Longyin Wen, Tiejian Luo, Yanjun Wu, and Pengfei Zhu. 2020. Spatial Attention Pyramid Network for Unsupervised Domain Adaptation. In European Conference on Computer Vision . 481–497

    2020

  27. [36]

    Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al

  28. [37]

    Shuai Li, Jianqiang Huang, Xian-Sheng Hua, and Lei Zhang. 2021. Category dictio- nary guided unsupervised domain adaptation for object detection. In Proceedings of the AAAI conference on artificial intelligence , Vol. 35. 1949–1957

    2021

  29. [38]

    Shuaifeng Li, Mao Ye, Xiatian Zhu, Lihua Zhou, and Lin Xiong. 2022. Source- free object detection by learning to overlook domain style. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 8014–8023

    2022

  30. [39]

    In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 10965–10975

  31. [40]

    Wuyang Li, Xinyu Liu, Xiwen Yao, and Yixuan Yuan. 2022. Scan: Cross domain object detection with semantic conditioned adaptation. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 36. 1421–1428

    2022

  32. [41]

    Wuyang Li, Xinyu Liu, and Yixuan Yuan. 2022. Sigma: Semantic-complete graph matching for domain adaptive object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 5291–5300

    2022

  33. [43]

    Xianfeng Li, Weijie Chen, Di Xie, Shicai Yang, Peng Yuan, Shiliang Pu, and Yueting Zhuang. 2021. A free lunch for unsupervised domain adaptive object detection without source data. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 35. 8474–8481

    2021

  34. [44]

    Yu-Jhe Li, Xiaoliang Dai, Chih-Yao Ma, Yen-Cheng Liu, Kan Chen, Bichen Wu, Zijian He, Kris Kitani, and Peter Vajda. 2022. Cross-domain adaptive teacher for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 7581–7590

    2022

  35. [45]

    Wuyang Li, Xinyu Liu, and Yixuan Yuan. 2023. Sigma++: Improved semantic- complete graph matching for domain adaptive object detection.IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)

    2023

  36. [46]

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. 2014. Microsoft coco: Common objects in context. In Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proce...

    2014

  37. [47]

    Dongnan Liu, Chaoyi Zhang, Yang Song, Heng Huang, Chenyu Wang, Michael Barnett, and Weidong Cai. 2022. Decompose to adapt: Cross-domain object detection via feature disentanglement. IEEE Transactions on Multimedia 25 (2022), 1333–1344

    2022

  38. [48]

    Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. 2017. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision . 2980–2988. MM ’24, October 28-November 1, 2024, Melbourne, VIC, Australia Boyong He, Yu...

    2017

  39. [49]

    Xinyu Liu, Wuyang Li, Qiushi Yang, Baopu Li, and Yixuan Yuan. 2022. Towards robust adaptive object detection under noisy annotations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 14207–14216

    2022

  40. [50]

    Yabo Liu, Jinghua Wang, Chao Huang, Yaowei Wang, and Yong Xu. 2023. CIGAR: Cross-Modality Graph Reasoning for Domain Adaptive Object Detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 23776–23786

    2023

  41. [51]

    Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. 2016. Ssd: Single shot multibox detec- tor. In Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part...

    2016

  42. [52]

    Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. 2022. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 11976–11986

    2022

  43. [53]

    Grace Luo, Lisa Dunlap, Dong Huk Park, Aleksander Holynski, and Trevor Darrell

  44. [54]

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. 2021. Swin transformer: Hierarchical vision transformer us- ing shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision. 10012–10022

    2021

  45. [55]

    Siyuan Qiao, Liang-Chieh Chen, and Alan Yuille. 2021. Detectors: Detecting objects with recursive feature pyramid and switchable atrous convolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 10213–10224

    2021

  46. [56]

    Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. 2021. Zero-shot text-to-image generation. In International conference on machine learning . Pmlr, 8821–8831

    2021

  47. [57]

    Joseph Redmon and Ali Farhadi. 2018. Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767 (2018)

    2018 arXiv

  48. [58]

    Muhammad Akhtar Munir, Muhammad Haris Khan, M Sarfraz, and Mohsen Ali

  49. [59]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 10684–10695

    2022

  50. [60]

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. 2015. U-net: Convolu- tional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceeding...

    2015

  51. [61]

    Aruni RoyChowdhury, Prithvijit Chakrabarty, Ashish Singh, SouYoung Jin, Huaizu Jiang, Liangliang Cao, and Erik Learned-Miller. 2019. Automatic adapta- tion of object detectors to new domains using self-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and ...

    2019

  52. [62]

    Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al

  53. [63]

    Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. 2016. Faster R-CNN: To- wards real-time object detection with region proposal networks.IEEE transactions on pattern analysis and machine intelligence 39, 6 (2016), 1137–1149

    2016

  54. [64]

    Kuniaki Saito, Yoshitaka Ushiku, Tatsuya Harada, and Kate Saenko. 2019. Strong- weak distribution alignment for adaptive object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 6956–6965

    2019

  55. [65]

    Christos Sakaridis, Dengxin Dai, and Luc Van Gool. 2018. Semantic foggy scene understanding with synthetic data. International Journal of Computer Vision 126 (2018), 973–992

    2018

  56. [66]

    Karen Simonyan and Andrew Zisserman. 2014. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)

    2014 arXiv

  57. [67]

    Jiaming Song, Chenlin Meng, and Stefano Ermon. 2020. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)

    2020 arXiv

  58. [68]

    Peng Su, Kun Wang, Xingyu Zeng, Shixiang Tang, Dapeng Chen, Di Qiu, and Xiaogang Wang. 2020. Adapting object detectors with conditional domain nor- malization. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XI 16 . Sp...

    2020

  59. [69]

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. 2022. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural infor...

    2022

  60. [70]

    Renshuai Tao, Hainan Li, Tianbo Wang, Yanlu Wei, Yifu Ding, Bowei Jin, Hong- ping Zhi, Xianglong Liu, and Aishan Liu. 2022. Exploring endogenous shift for cross-domain detection: A large-scale benchmark and perturbation suppression network. In 2022 IEEE/CVF Conference on Compu...

    2022

  61. [71]

    Antti Tarvainen and Harri Valpola. 2017. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems 30 (2017)

    2017

  62. [72]

    Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. 2020. FCOS: A simple and strong anchor-free object detector. IEEE Transactions on Pattern Analysis and Machine Intelligence 44, 4 (2020), 1922–1933

    2020

  63. [73]

    Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. 2023. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 1921–1930

    2023

  64. [74]

    Chien-Yao Wang, Alexey Bochkovskiy, and Hong-Yuan Mark Liao. 2023. YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recogni- tion. 7464–7475

    2023

  65. [75]

    Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. 2023. Emergent correspondence from image diffusion. Advances in Neural Information Processing Systems 36 (2023), 1363–1389

    2023

  66. [76]

    Aming Wu, Rui Liu, Yahong Han, Linchao Zhu, and Yi Yang. 2021. Vector- decomposed disentanglement for domain-invariant object detection. In Proceed- ings of the IEEE/CVF International Conference on Computer Vision . 9342–9351

    2021

  67. [77]

    Chang-Dong Xu, Xing-Ran Zhao, Xin Jin, and Xiu-Shen Wei. 2020. Exploring categorical regularization for domain adaptive object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 11724– 11733

    2020

  68. [78]

    Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. 2023. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2955–2966

    2023

  69. [79]

    Jayeon Yoo, Inseop Chung, and Nojun Kwak. 2022. Unsupervised domain adap- tation for one-stage object detector using offsets to bounding box. In European Conference on Computer Vision . Springer, 691–708

    2022

  70. [80]

    Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. 2020. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognitio...

    2020

  71. [81]

    Aming Wu, Yahong Han, Linchao Zhu, and Yi Yang. 2021. Instance-invariant domain adaptive object detection via progressive disentanglement. IEEE Trans- actions on Pattern Analysis and Machine Intelligence 44, 8 (2021), 4178–4193

    2021

  72. [82]

    Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Heung-Yeung Shum. 2022. DINO: DETR with Improved DeNoising Anchor Boxes for End-to-End Object Detection. In The Eleventh International Conference on Learning Representations

    2022

  73. [83]

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. 2023. Adding conditional con- trol to text-to-image diffusion models. InProceedings of the IEEE/CVF International Conference on Computer Vision . 3836–3847

    2023

  74. [84]

    Shifeng Zhang, Cheng Chi, Yongqiang Yao, Zhen Lei, and Stan Z Li. 2020. Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 9759–9768

    2020

  75. [85]

    Liang Zhao and Limin Wang. 2022. Task-specific inconsistency alignment for domain adaptive object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 14217–14226

    2022

  76. [87]

    Jinze Yu, Jiaming Liu, Xiaobao Wei, Haoyi Zhou, Yohei Nakata, Denis Gudovskiy, Tomoyuki Okuno, Jianxin Li, Kurt Keutzer, and Shanghang Zhang. 2022. MT- Trans: Cross-domain object detection with mean teacher transformer. InEuropean Conference on Computer Vision . Springer, 629–645

    2022

  77. [88]

    Wenzhang Zhou, Dawei Du, Libo Zhang, Tiejian Luo, and Yanjun Wu. 2022. Multi- granularity alignment domain adaptation for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 9581–9590

    2022

  78. [89]

    Wenzhang Zhou, Heng Fan, Tiejian Luo, and Libo Zhang. 2023. Unsupervised Domain Adaptive Detection with Network Stability Analysis. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 6986–6995. Diffusion Domain Teacher: Diffusion Guided Domain Adaptiv...

    2023

  79. [90]

    Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. 2017. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Pro- ceedings of the IEEE international conference on computer vision . 2223–2232

    2017

  80. [91]

    Xinge Zhu, Jiangmiao Pang, Ceyuan Yang, Jianping Shi, and Dahua Lin. 2019. Adapting object detectors via selective cross-domain alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 687–696

    2019

  81. [92]

    Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. 2021. Deformable DETR: Deformable Transformers for End-to-End Object Detection. In International Conference on Learning Representations . MM ’24, October 28-November 1, 2024, Melbourne, VIC, Australia Boyo...

    2021

  82. [93]

    Zhen Zhao, Yuhong Guo, Haifeng Shen, and Jieping Ye. 2020. Adaptive object detection with dual multi-label prediction. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXVIII

    2020

  83. [99]

    FoggyCityscapes [65] is a dataset where foggy weather conditions are added to Cityscapes, with similar images and identical labels from Cityscapes

    (50.3 mAP) and HT [15] (50.4 mAP). FoggyCityscapes [65] is a dataset where foggy weather conditions are added to Cityscapes, with similar images and identical labels from Cityscapes. We ob- serve that our DDT method does not demonstrate superior perfor- mance in inter-domain a...

    2024

  84. [2015]

    International journal of computer vision 115 (2015), 211–252

    Imagenet large scale visual recognition challenge. International journal of computer vision 115 (2015), 211–252

    2015

  85. [2019]

    InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Diversify and match: A domain adaptive representation learning paradigm for object detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 12456–12465

  86. [2021]

    Advances in Neural Information Processing Systems 34 (2021), 22770–22782

    Ssal: Synergizing between self-training and adversarial learning for domain adaptive object detection. Advances in Neural Information Processing Systems 34 (2021), 22770–22782

    2021

  87. [2022]

    In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 16000–16009

  88. [2024]

    Advances in Neural Information Processing Systems 36 (2024)

    Diffusion hyperfeatures: Searching through time and space for semantic correspondence. Advances in Neural Information Processing Systems 36 (2024)

    2024

Pith tools

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