Pith. sign in

REVIEW 4 major objections 5 minor 47 references

Universal Domain Adaptation for Semantic Segmentation

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper introduces UniDA-SS, a setting where a segmentation model adapts to an unlabeled target domain without knowing which classes are shared, and proposes UniMAP, which separates common from private pixels using source/target…

desk verdict The new UniDA-SS task and UniMAP method are interesting, but the benchmark's class overlap is internally inconsistent and the SOTA comparisons are confounded, so the headline numbers should not be trusted as published. read the letter →

arxiv 2505.22458 v2 pith:OIS7DJRB submitted 2025-05-28 cs.CV

Seun-An Choe , Keon-Hee Park , Jinwoo Choi , Gyeong-Moon Park This is my paper
classification cs.CV
keywords universaldomainadaptationsemanticsegmentationopenpartialprototypelearningself-trainingpseudo-labelweightingunknownclassdetectiondomain-specificprototypes
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

This paper introduces a new problem setting, Universal Domain Adaptation for Semantic Segmentation (UniDA-SS), in which a segmentation model must transfer from labeled source data to unlabeled target data without being told which classes the two domains share. The paper argues that existing UDA methods degrade when source-private or target-private classes are present, and proposes UniMAP to keep learning common classes while labeling target-private pixels as unknown. The core idea is to give every class two prototypes, one for the source domain and one for the target domain, so that common-class pixel embeddings become roughly equidistant to both while private-class embeddings stay close to only one; a pixel weight built from those two distances then up-weights likely-common pixels. A second component, Target-based Image Matching, pairs each target image with the source image whose class content best matches the target pseudo-label, prioritizing rare common classes. If the paper is right, pixel-level domain adaptation can succeed with no prior category-overlap knowledge, reaching H-scores of 41.33 on Pascal-Context to Cityscapes and 45.51 on GTA5 to IDD.

What carries the argument

The load-bearing object is the pair of domain-specific prototypes per class, held in a fixed Simplex Equiangular Tight Frame (ETF) so that all prototype pairs share the same cosine similarity and L2 norm. Domain-Specific Prototype-based Distinction (DSPD) uses these prototypes with three losses — cross-entropy, pixel-prototype contrastive, and pixel-prototype distance optimization — to shape the embedding space; the pixel-wise weight $w$ computed from the pixel's cosine similarity to the source and target prototypes of its class turns prototype geometry into a common/private classifier. Target-based Image Matching (TIM) is the second mechanism: it estimates the target pseudo-label's class distribution, reweights rare classes, scores source images by overlap with that distribution, and pairs the best source image with the target image in the batch. Together, the two mechanisms separate common from private pixels and concentrate training signal on common classes.

What would settle it

On a held-out set of target pixels with known labels, compute the weight $w$ from Eq. (10) for each pixel and measure how well it separates common-class from private-class pixels (for example, by area under the ROC curve); if the separation is near chance, the geometric assumption behind DSPD's weight scaling is not doing the work the paper attributes to it.

Watch

Extended reading notes

Core claim

The paper's central claim is that the unknown-overlap problem in UDA semantic segmentation can be solved by modeling each class with two domain-specific prototypes rather than one shared prototype. UniMAP places a source and a target prototype for every class in a fixed simplex equiangular tight frame, trains with prototype-based losses so that common-class pixel embeddings become similar to both prototypes, and defines a pixel weight $w = 2(d_s+1)(d_t+1)/((d_s+1)+(d_t+1))$ from the two cosine similarities. Because common-class embeddings are close to both prototypes and private-class embeddings are close to only one, $w$ is larger for common-class pixels; applying $w$ to the pseudo-label confidence and to the target loss prevents common classes from being misassigned as unknown. Target-based Image Matching then pairs each target image with the labeled source image whose class distribution overlaps most under a rare-class-weighted similarity score, increasing the presence of common classes in every batch. The paper reports that this combination improves H-score from 30.11 to 41.33 on Pascal-Context to Cityscapes and from 41.26 to 45.51 on GTA5 to IDD, while remaining competitive across closed-set, open-set, partial, and open-partial settings.

Load-bearing premise

The method assumes that after training, common-class pixel embeddings become roughly equally similar to the source and target prototypes of their class, while private-class embeddings are similar to only one prototype; the weight formula and the common/private separation rest on this geometric claim, which the paper does not prove.

Editorial extensions

If this is right

  • A deployed segmentation model can adapt to a new, unlabeled environment without the user specifying which categories exist there; target-private regions are reported as unknown.
  • Common-class pixels that previously fell below the pseudo-label confidence threshold and were absorbed into the unknown class are retained, which is the direct mechanism behind the improved common-class scores.
  • Rare common classes receive more training signal because TIM's source-image selection is weighted toward classes that are rare in the target pseudo-label.
  • The method raises H-score by about 11 points on Pascal-Context to Cityscapes and about 4 points on GTA5 to IDD over the strongest prior baseline, with the larger gains in target-private detection.
  • The same framework generalizes across closed-set, open-set, partial, and open-partial settings, so a practitioner does not need to identify which adaptation scenario they are in.

Reading between the lines

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

  • The two-prototype geometry is not tied to exactly two domains; assigning each class one prototype per source dataset or per environmental condition inside an ETF-like frame could extend UniMAP to multi-source and continual adaptation, though the closed-form weight would need a multi-way generalization.
  • The weight $w$ is a harmonic-mean-like combination of the two cosine similarities; a testable extension is to replace it with a small learned network on $(d_s, d_t)$, which could capture correlations the fixed formula misses.
  • TIM already estimates category overlap from target pseudo-labels; that estimate could be output as a by-product and used to calibrate the unknown threshold per class, a step the paper leaves implicit.
  • The new benchmarks' overlap splits are part of the contribution; documenting the source-label filtering for the private classes would make the reported H-scores independently reproducible and testable against alternative overlap definitions.
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 introduces Universal Domain Adaptation for Semantic Segmentation (UniDA-SS), a setup in which source and target domains may contain both source-private and target-private classes and the overlap is unknown. The authors propose UniMAP, which combines Domain-Specific Prototype-based Distinction (DSPD), using two prototypes per class in a fixed ETF space with a pixel-wise weight scaling factor, and Target-based Image Matching (TIM), which selects source images with high common-class overlap with target pseudo-labels. They define two new OPDA-SS benchmarks, Pascal-Context to Cityscapes and GTA5 to IDD, and report state-of-the-art H-scores (41.33 and 45.51) over prior UDA-SS and UniDA baselines. The paper includes extensive ablations showing that each component contributes positively when combined.

Significance. If the empirical results are correct and reproducible, the paper makes a useful contribution: it is among the first to address universal adaptation for semantic segmentation, proposes a modular method with a released codebase, and provides component-wise ablations. The dual-prototype idea and the image-matching strategy are plausible and could be of interest to the UDA community. However, the significance is currently contingent on the validity of the newly proposed benchmarks, which are used for all headline numbers. The reported improvements over BUS are substantial, but the benchmark protocol as written is underspecified and appears to conflict with the standard label sets of the underlying datasets. The code release is a positive step, but the evaluation confounds and missing reproducibility details prevent the current version from supporting the SOTA claims.

major comments (4)
  1. [Section 4.1 (Pascal-Context to Cityscapes benchmark)] The paper lists seven target-private classes for Pascal-Context to Cityscapes: pole, light, sign, terrain, person, rider, and train. However, Pascal-Context contains annotations for person and train, and likely also for rider under its standard 59-class label set. Under any natural label mapping, these should be common classes rather than target-private. The paper does not specify the 12 common classes, nor does it describe any procedure for filtering the source labels to remove person, train, and related classes. Without this mapping, the Private IoU and H-score are not well-defined, and the results in Table 1 cannot be reproduced or compared across methods.
  2. [Section 4.1 (GTA5 to IDD benchmark)] The paper states that terrain and train are source-private classes in GTA5 to IDD, but both GTA5 and IDD include terrain and train in their standard taxonomies. This suggests either a nonstandard label mapping or an incomplete description of the protocol. No source-mask filtering or class-removal procedure is described, so the claimed private-class evaluation for this benchmark is ambiguous and potentially invalid.
  3. [Section 4.1, Baselines paragraph] The UniDA classification baselines (UAN, UniOT, MLNet) are evaluated with a DeepLabv2 network and ResNet-101 backbone, while the UDA-SS baselines (DAFormer, HRDA, MIC, BUS) and UniMAP use a MiT-B5 encoder. The large performance gaps for the UniDA baselines in Tables 1 and 2 are likely dominated by this architecture mismatch rather than by algorithmic ability. For a fair SOTA comparison, report all methods with the same backbone, or clearly state that the UniDA baselines are not directly comparable.
  4. [Section 3.3, Eq. (10) and Table 4] The weight scaling factor w relies on the geometric assumption that common-class pixel embeddings become similarly close to the source and target prototypes, while private-class embeddings are close to only one prototype. This assumption is not empirically validated. The ablation in Table 4 shows that applying w alone reduces the H-score from 36.03 to 31.08, so the weighting is not reliable on its own. Please provide direct evidence for the distance-ratio assumption (e.g., distributions of ds and dt for common versus private pixels) and explain why w alone degrades performance.
minor comments (5)
  1. [Section 3.3, Eq. (5)] The constant r in the ETF definition is not defined; please specify its value and the dimension of the prototype space.
  2. [Section 3.4, TIM] The text says the selected source image is paired with 'the corresponding target image' in a batch, but the implementation uses a batch of two random crops. Clarify how a single target image is selected and how this interacts with the DACS mixing strategy.
  3. [Section 4.1, Implementation Details] There is a typo: 'muli-resolution' should be 'multi-resolution'.
  4. [Tables 1 and 2] No error bars or multiple-seed results are reported; please include at least three random seeds or explicitly state that results are from a single run.
  5. [Throughout, Eqs. (5), (11), (12)] The symbol C is used without a clear definition; clarify whether C denotes the number of source classes, the number of common classes, or something else, since Eqs. (1) and (11) appear to use different class counts.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation: UniMAP's components are heuristic and empirically evaluated, not fitted to the target H-score; the only self-citation (BUS baseline) is legitimate prior work and non-load-bearing.

full rationale

I found no circular step in the paper's derivation chain. The proposed components (DSPD, TIM, and the weighting w in Eq. 10) are heuristic mechanisms trained through losses (Eqs. 6-9, 11-12) and evaluated on external benchmarks; they do not fit or directly predict the reported H-scores or Private IoUs. The baseline is built on the authors' own BUS [6], but BUS is a published, externally comparable method, and Section 3.2 explicitly describes which BUS components are removed, so this is a legitimate baseline rather than a self-fulfilling construction. The geometric assumption that common-class embeddings align with both source and target prototypes is unproven, but that is a correctness/robustness concern, not circularity. One non-circular issue worth flagging is the under-specified benchmark protocol: Section 4.1 states that for Pascal-Context→Cityscapes the 7 target-private classes are 'pole', 'light', 'sign', 'terrain', 'person', 'rider', and 'train', while Pascal-Context annotations include person and train; the paper does not state how source labels are filtered or which 12 classes are treated as common. This makes the Private IoU and H-score comparisons harder to verify across methods, but it is an evaluation-protocol ambiguity rather than a circular derivation, so it does not increase the circularity score.

Assumptions & free parameters 6 free parameters · 4 assumptions · 1 invented entities

The central mechanism rests on a geometric alignment heuristic, reliable teacher pseudo-labels, and hand-set hyperparameters. The paper does not prove these conditions hold, and the benchmark construction adds a further assumption about how source labels are filtered.

free parameters (6)
  • tau_p = 0.5
    Threshold for assigning unknown pseudo-labels in Eq. 2 and Eq. 12; chosen by hand and directly controls the common/unknown tradeoff.
  • tau_t = 0.968
    Image-level reliability threshold in Eq. 3 used to weight the target segmentation loss.
  • lambda_1 = 0.01
    Weight for the pixel-prototype contrastive loss L_PPC in Eq. 9.
  • lambda_2 = 0.01
    Weight for the pixel-prototype distance loss L_PPD in Eq. 9.
  • tau = 0.1
    Temperature in the pixel-prototype contrastive loss in Eq. 7.
  • T = 0.01
    Temperature in the TIM class-frequency weighting in Eq. 14.
assumptions (4)
  • domain assumption Fixed ETF prototype geometry gives stable source-target prototype separation.
    Eq. 5 assigns 2C+1 prototypes from a simplex ETF; the paper assumes this fixed geometry provides equal cosine separation without learning prototype positions.
  • ad hoc to paper Common-class pixels become similarly close to source and target prototypes, while private pixels are close to only one.
    This geometric claim is stated in Section 3.3 and used in Eq. 10 and Eq. 12, but it is not derived or independently validated.
  • domain assumption Teacher EMA pseudo-labels are reliable enough to define target class distributions and source image pairing.
    Eqs. 2, 3, and 13-15 all depend on pseudo-label quality; if pseudo-labels are inaccurate, TIM selects the wrong source images and weights the wrong pixels.
  • ad hoc to paper The manually defined benchmarks accurately instantiate an unknown category-shift setting.
    Section 4.1 manually designates seven Cityscapes classes as target-private in Pascal-Context to Cityscapes without stating how source labels for those classes are handled.
invented entities (1)
  • Domain-specific dual prototypes in a fixed ETF space
    purpose: Separate domain-specific features per class and identify common classes via relative distances to source and target prototypes.
    The construct is evaluated only through the paper's benchmarks and ablations; no external falsifiable handle is provided outside this paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Universal Domain Adaptation for Semantic Segmentation." pith.science (2026). https://pith.science/paper/OIS7DJRB

@misc{pith2026250522458,
  author       = {Pith},
  title        = {Pith review of: Universal Domain Adaptation for Semantic Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OIS7DJRB}},
  note         = {Machine review of arXiv:2505.22458}
}
read the original abstract

Unsupervised domain adaptation for semantic segmentation (UDA-SS) aims to transfer knowledge from labeled source data to unlabeled target data. However, traditional UDA-SS methods assume that category settings between source and target domains are known, which is unrealistic in real-world scenarios. This leads to performance degradation if private classes exist. To address this limitation, we propose Universal Domain Adaptation for Semantic Segmentation (UniDA-SS), achieving robust adaptation even without prior knowledge of category settings. We define the problem in the UniDA-SS scenario as low confidence scores of common classes in the target domain, which leads to confusion with private classes. To solve this problem, we propose UniMAP: UniDA-SS with Image Matching and Prototype-based Distinction, a novel framework composed of two key components. First, Domain-Specific Prototype-based Distinction (DSPD) divides each class into two domain-specific prototypes, enabling finer separation of domain-specific features and enhancing the identification of common classes across domains. Second, Target-based Image Matching (TIM) selects a source image containing the most common-class pixels based on the target pseudo-label and pairs it in a batch to promote effective learning of common classes. We also introduce a new UniDA-SS benchmark and demonstrate through various experiments that UniMAP significantly outperforms baselines. The code is available at https://github.com/KU-VGI/UniMAP.

Figures

Figures reproduced from arXiv: 2505.22458 by the authors.

Figure 1
Figure 1. Visualization results of the UDA-SS models across dif [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of our proposed method, UniMAP. The top right illustrates the main training framework. The model is optimized with [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Qualitative results of OPDA-SS setting. We visualize the segmentation predictions from different methods on the Cityscapes [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 39 canonical work pages

  1. [1]

    On the effectiveness of image rotation for open set domain adaptation

    Silvia Bucci, Mohammad Reza Loghmani, and Tatiana Tom- masi. On the effectiveness of image rotation for open set domain adaptation. InEuropean conference on computer vi- sion, pages 422–438. Springer, 2020. 3

    work page 2020

  2. [2]

    Unified optimal transport framework for universal domain adaptation.Advances in Neural Information Processing Sys- tems, 35:29512–29524, 2022

    Wanxing Chang, Ye Shi, Hoang Tuan, and Jingya Wang. Unified optimal transport framework for universal domain adaptation.Advances in Neural Information Processing Sys- tems, 35:29512–29524, 2022. 6

    work page 2022

  3. [3]

    Synergistic image and feature adaptation: Towards cross-modality domain adaptation for medical image seg- mentation

    Cheng Chen, Qi Dou, Hao Chen, Jing Qin, and Pheng-Ann Heng. Synergistic image and feature adaptation: Towards cross-modality domain adaptation for medical image seg- mentation. InProceedings of the AAAI conference on ar- tificial intelligence, pages 865–872, 2019. 3

    work page 2019

  4. [4]

    Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolu- tion, and fully connected crfs.IEEE transactions on pattern analysis and machine intelligence, 40(4):834–848, 2017. 3, 6

    work page 2017

  5. [5]

    Do- main adaptation for semantic segmentation with maximum squares loss

    Minghao Chen, Hongyang Xue, and Deng Cai. Do- main adaptation for semantic segmentation with maximum squares loss. InProceedings of the IEEE/CVF international conference on computer vision, pages 2090–2099, 2019. 3

    work page 2019

  6. [6]

    Open-set domain adaptation for semantic segmentation

    Seun-An Choe, Ah-Hyung Shin, Keon-Hee Park, Jinwoo Choi, and Gyeong-Moon Park. Open-set domain adaptation for semantic segmentation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23943–23953, 2024. 2, 3, 6, 7

    work page 2024

  7. [7]

    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. InProceed- ings of the IEEE conference on computer vision and pattern recognition, pages 3213–3223, 2016. 6

    2016

  8. [8]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee, 2009. 6

    2009

Show all 47 references
  1. [9]

    Ssf-dan: Separated semantic feature based domain adaptation network for semantic segmentation

    Liang Du, Jingang Tan, Hongye Yang, Jianfeng Feng, Xi- angyang Xue, Qibao Zheng, Xiaoqing Ye, and Xiaolin Zhang. Ssf-dan: Separated semantic feature based domain adaptation network for semantic segmentation. InProceed- ings of the IEEE/CVF International Conference on Com- pute...

    2019

  2. [10]

    Learning to detect open classes for universal domain adapta- tion

    Bo Fu, Zhangjie Cao, Mingsheng Long, and Jianmin Wang. Learning to detect open classes for universal domain adapta- tion. InComputer Vision–ECCV 2020: 16th European Con- ference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XV 16, pages 567–583. Springer, 2020. 3

    2020

  3. [11]

    Deep residual learning for image recognition

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

    2016

  4. [12]

    Conditional generative adversarial network for struc- tured domain adaptation

    Weixiang Hong, Zhenzhen Wang, Ming Yang, and Junsong Yuan. Conditional generative adversarial network for struc- tured domain adaptation. InProceedings of the IEEE con- ference on computer vision and pattern recognition, pages 1335–1344, 2018. 3

    2018

  5. [13]

    Daformer: Improving network architectures and training strategies for domain-adaptive semantic segmentation

    Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Daformer: Improving network architectures and training strategies for domain-adaptive semantic segmentation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9924–9935, 2022. 3, 6

    2022

  6. [14]

    Hrda: Context-aware high-resolution domain-adaptive semantic segmentation

    Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Hrda: Context-aware high-resolution domain-adaptive semantic segmentation. InEuropean Conference on Computer Vision, pages 372–391. Springer, 2022. 3, 6

    2022

  7. [15]

    Mic: Masked image consistency for context- enhanced domain adaptation

    Lukas Hoyer, Dengxin Dai, Haoran Wang, and Luc Van Gool. Mic: Masked image consistency for context- enhanced domain adaptation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11721–11732, 2023. 2, 6

    2023

  8. [16]

    Learning texture invari- ant representation for domain adaptation of semantic seg- mentation

    Myeongjin Kim and Hyeran Byun. Learning texture invari- ant representation for domain adaptation of semantic seg- mentation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12975– 12984, 2020. 3

    2020

  9. [17]

    Improv- ing semantic segmentation via decoupled body and edge su- pervision

    Xiangtai Li, Xia Li, Li Zhang, Guangliang Cheng, Jianping Shi, Zhouchen Lin, Shaohua Tan, and Yunhai Tong. Improv- ing semantic segmentation via decoupled body and edge su- pervision. InComputer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proce...

    2020

  10. [18]

    Bidirectional learning for domain adaptation of semantic segmentation

    Yunsheng Li, Lu Yuan, and Nuno Vasconcelos. Bidirectional learning for domain adaptation of semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 6936–6945, 2019. 3

    2019

  11. [19]

    Constructing self-motivated pyramid curriculums for cross- domain semantic segmentation: A non-adversarial approach

    Qing Lian, Fengmao Lv, Lixin Duan, and Boqing Gong. Constructing self-motivated pyramid curriculums for cross- domain semantic segmentation: A non-adversarial approach. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 6758–6767, 2019. 3

    2019

  12. [20]

    Parsenet: Looking wider to see better.arXiv preprint arXiv:1506.04579, 2015

    Wei Liu, Andrew Rabinovich, and Alexander C Berg. Parsenet: Looking wider to see better.arXiv preprint arXiv:1506.04579, 2015. 3

    2015 arXiv

  13. [21]

    Fully convolutional networks for semantic segmentation

    Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. InPro- ceedings of the IEEE conference on computer vision and pat- tern recognition, pages 3431–3440, 2015. 3

    2015

  14. [22]

    Mlnet: Mutual learning network with neighbor- hood invariance for universal domain adaptation

    Yanzuo Lu, Meng Shen, Andy J Ma, Xiaohua Xie, and Jian- Huang Lai. Mlnet: Mutual learning network with neighbor- hood invariance for universal domain adaptation. InPro- ceedings of the AAAI Conference on Artificial Intelligence, pages 3900–3908, 2024. 6

    2024

  15. [23]

    Pixmatch: Unsu- pervised domain adaptation via pixelwise consistency train- ing

    Luke Melas-Kyriazi and Arjun K Manrai. Pixmatch: Unsu- pervised domain adaptation via pixelwise consistency train- ing. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 12435–12445,

  16. [24]

    The role of context for object detection and semantic segmentation in the wild

    Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. InProceedings of the IEEE conference on computer vision and pattern recogni-...

    2014

  17. [25]

    Unsupervised intra-domain adaptation for se- mantic segmentation through self-supervision

    Fei Pan, Inkyu Shin, Francois Rameau, Seokju Lee, and In So Kweon. Unsupervised intra-domain adaptation for se- mantic segmentation through self-supervision. InProceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3764–3773, 2020. 3

    2020

  18. [26]

    Prevalence of neural collapse during the terminal phase of deep learning training.Proceedings of the National Academy of Sciences, 117(40):24652–24663, 2020

    Vardan Papyan, XY Han, and David L Donoho. Prevalence of neural collapse during the terminal phase of deep learning training.Proceedings of the National Academy of Sciences, 117(40):24652–24663, 2020. 4

    2020

  19. [27]

    Playing for data: Ground truth from computer games

    Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. InComputer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 102–118. Springer,

    2016

  20. [28]

    Ovanet: One-vs-all net- work for universal domain adaptation

    Kuniaki Saito and Kate Saenko. Ovanet: One-vs-all net- work for universal domain adaptation. InProceedings of the ieee/cvf international conference on computer vision, pages 9000–9009, 2021. 3

    2021

  21. [29]

    Universal domain adaptation through self supervi- sion.Advances in neural information processing systems, 33:16282–16292, 2020

    Kuniaki Saito, Donghyun Kim, Stan Sclaroff, and Kate Saenko. Universal domain adaptation through self supervi- sion.Advances in neural information processing systems, 33:16282–16292, 2020. 3

    2020

  22. [30]

    Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results.Advances in neural information processing systems, 30, 2017

    Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results.Advances in neural information processing systems, 30, 2017. 3

    2017

  23. [31]

    Dacs: Domain adaptation via cross- domain mixed sampling

    Wilhelm Tranheden, Viktor Olsson, Juliano Pinto, and Lennart Svensson. Dacs: Domain adaptation via cross- domain mixed sampling. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1379–1389, 2021. 3, 4, 6

    2021

  24. [32]

    Learning to adapt structured output space for semantic seg- mentation

    Yi-Hsuan Tsai, Wei-Chih Hung, Samuel Schulter, Ki- hyuk Sohn, Ming-Hsuan Yang, and Manmohan Chandraker. Learning to adapt structured output space for semantic seg- mentation. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 7472–7481,

  25. [33]

    Domain adaptation for structured output via discriminative patch representations

    Yi-Hsuan Tsai, Kihyuk Sohn, Samuel Schulter, and Manmo- han Chandraker. Domain adaptation for structured output via discriminative patch representations. InProceedings of the IEEE/CVF international conference on computer vision, pages 1456–1465, 2019. 3

    2019

  26. [34]

    Idd: A dataset for exploring problems of autonomous navigation in uncon- strained environments

    Girish Varma, Anbumani Subramanian, Anoop Namboodiri, Manmohan Chandraker, and CV Jawahar. Idd: A dataset for exploring problems of autonomous navigation in uncon- strained environments. In2019 IEEE winter conference on applications of computer vision (WACV), pages 1743–1751. ...

    2019

  27. [35]

    Domain adaptive semantic segmentation with self-supervised depth estimation

    Qin Wang, Dengxin Dai, Lukas Hoyer, Luc Van Gool, and Olga Fink. Domain adaptive semantic segmentation with self-supervised depth estimation. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 8515–8525, 2021. 3

    2021

  28. [36]

    Uncertainty-aware pseudo label refinery for domain adaptive semantic segmentation

    Yuxi Wang, Junran Peng, and ZhaoXiang Zhang. Uncertainty-aware pseudo label refinery for domain adaptive semantic segmentation. InProceedings of the IEEE/CVF international conference on computer vision, pages 9092–9101, 2021. 3

    2021

  29. [37]

    Segformer: Simple and efficient design for semantic segmentation with transform- ers.Advances in Neural Information Processing Systems, 34:12077–12090, 2021

    Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transform- ers.Advances in Neural Information Processing Systems, 34:12077–12090, 2021. 3, 6

    2021

  30. [38]

    Universal domain adaptation

    Kaichao You, Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Universal domain adaptation. InProceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 2720–2729, 2019. 3, 6

    2019

  31. [39]

    Faster segment anything: Towards lightweight sam for mo- bile applications.arXiv preprint arXiv:2306.14289, 2023

    Chaoning Zhang, Dongshen Han, Yu Qiao, Jung Uk Kim, Sung-Ho Bae, Seungkyu Lee, and Choong Seon Hong. Faster segment anything: Towards lightweight sam for mo- bile applications.arXiv preprint arXiv:2306.14289, 2023. 6

    2023 arXiv

  32. [40]

    Prototypical pseudo label denoising and tar- get structure learning for domain adaptive semantic segmen- tation

    Pan Zhang, Bo Zhang, Ting Zhang, Dong Chen, Yong Wang, and Fang Wen. Prototypical pseudo label denoising and tar- get structure learning for domain adaptive semantic segmen- tation. InProceedings of the IEEE/CVF conference on com- puter vision and pattern recognition, pages 12...

  33. [41]

    Pyramid scene parsing network

    Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2881–2890, 2017. 3

    2017

  34. [42]

    Psanet: Point- wise spatial attention network for scene parsing

    Hengshuang Zhao, Yi Zhang, Shu Liu, Jianping Shi, Chen Change Loy, Dahua Lin, and Jiaya Jia. Psanet: Point- wise spatial attention network for scene parsing. InProceed- ings of the European conference on computer vision (ECCV), pages 267–283, 2018. 3

    2018

  35. [43]

    Rethinking semantic segmen- tation from a sequence-to-sequence perspective with trans- formers

    Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmen- tation from a sequence-to-sequence perspective with trans- formers. InProceedings of the IEEE/CVF conference ...

  36. [44]

    Rethinking semantic segmentation: A proto- type view

    Tianfei Zhou, Wenguan Wang, Ender Konukoglu, and Luc Van Gool. Rethinking semantic segmentation: A proto- type view. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2582– 2593, 2022. 4

    2022

  37. [45]

    Asymmetric non-local neural networks for seman- tic segmentation

    Zhen Zhu, Mengde Xu, Song Bai, Tengteng Huang, and Xi- ang Bai. Asymmetric non-local neural networks for seman- tic segmentation. InProceedings of the IEEE/CVF Inter- national Conference on Computer Vision, pages 593–602,

  38. [46]

    Unsupervised domain adaptation for semantic segmentation via class-balanced self-training

    Yang Zou, Zhiding Yu, BVK Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. InProceedings of the Eu- ropean conference on computer vision (ECCV), pages 289– 305, 2018. 3

    2018

  39. [47]

    Confidence regularized self-training

    Yang Zou, Zhiding Yu, Xiaofeng Liu, BVK Kumar, and Jin- song Wang. Confidence regularized self-training. InPro- ceedings of the IEEE/CVF international conference on com- puter vision, pages 5982–5991, 2019. 3

    2019

Pith tools

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