Pith. sign in

REVIEW 4 major objections 6 minor 114 references

Prototypical Progressive Alignment and Reweighting for Generalizable Semantic Segmentation

T0 review · 4 major / 6 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 aligning a segmentation model's features to CLIP text prototypes, in an easy-to-difficult schedule, and down-weighting outlier pixels yields state-of-the-art generalization to unseen domains.

desk verdict A genuinely new CLIP-based prototype mechanism with strong empirical support; the main fix is specifying the alignment loss. read the letter →

arxiv 2507.11955 v1 pith:LPX6HSLC submitted 2025-07-16 cs.CV

Yuhang Zhang , Zhengyu Zhang , Muxin Liao , Shishun Tian , Wenbin Zou , Lu Zhang , Chen Xu This is my paper
classification cs.CV
keywords domaingeneralizationsemanticsegmentationprototypicalalignmentCLIPtextprototypespixelreweightingnegativetransferurbansceneunderstanding
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 aims to establish that a semantic segmentation model trained only on labeled source images can generalize to unseen target domains if its features are pulled toward prototypes derived from CLIP text embeddings instead of toward prototypes averaged from source data. The proposed PPAR framework builds two kinds of text prototype per class: an Original Text Prototype (OTP) from the class name alone, and a Visual Text Prototype (VTP) that appends the class's dominant color and local texture pattern to the class name. Shallow features are aligned to the easier VTP and deep features to the harder OTP, a progressive easy-to-difficult schedule, while a reweighting term down-weights pixels whose features have low similarity to the OTP, treating them as outliers. On five single-source generalization settings with a ResNet-50 backbone, the method reports an average mIoU of 43.1%, surpassing prior state-of-the-art methods, and it also improves results on lightweight and vision-foundation-model backbones. The reader should care because the method uses only frozen CLIP text features at training time and adds no cost at inference.

What carries the argument

The central machinery is a pair of CLIP text prototypes per class used as alignment anchors. The Original Text Prototype ($\mathrm{OTP}$) is the frozen CLIP text embedding of a class name, and the Visual Text Prototype ($\mathrm{VTP}$) is the CLIP text embedding of the class name concatenated with its dominant color and local-binary-pattern texture; both are generated once and not learned. The argument is carried by two KL-divergence alignment losses, shallow-to-VTP and deep-to-OTP, plus a pixel-level reweighting multiplier derived from similarity to the OTP. These components replace the naive batch-averaged prototype and equal-weight training of prior prototypical alignment methods.

What would settle it

Run PPAR on a source domain and evaluate on a held-out target far outside CLIP's training distribution, such as fisheye or aerial street scenes; if average mIoU no longer exceeds the ISW baseline, the claim that text-prototype similarity measures transferability is falsified. A second check: compute the KL alignment losses before and after removing the dimension-mapping to 512 and any normalization; if the losses can be minimized without improving generalization, the alignment term is not the cause of the gains.

Watch

Extended reading notes

Core claim

On its own terms, the paper's claim is that prototypical alignment for domain generalization fails when prototypes are computed as batch centroids of source features, because those centroids are source-biased and vulnerable to outliers. PPAR replaces them with CLIP-generated text prototypes: $\mathrm{OTP}$ is the CLIP text embedding of the class name, and $\mathrm{VTP}$ is the CLIP text embedding of the class name augmented with the most frequent color and local-binary-pattern texture string in the source class. Alignment is done progressively by minimizing KL divergence between shallow class centroids and the VTP (the easy, visual-information step) and between deep class centroids and the OTP (the difficult, category-only step). A prototypical reweighting loss multiplies each pixel's cross-entropy contribution by $\exp(-(U-U_{\min})/(U_{\max}-U_{\min}))$, where $U$ is the entropy of the pixel's softmax similarity to OTPs, so outlier-like source pixels contribute less. The paper reports state-of-the-art average mIoU of 43.1% on ResNet-50, 37.4% on ShuffleNet-V2 and 38.2% on MobileNet-V2, plus gains over prior methods when PPAR components are added to them. A theoretical section connects the design to a domain-generalization bound, arguing that reweighting reduces source risk while progressive alignment reduces inter-source divergence.

Load-bearing premise

The load-bearing assumption is that CLIP's text embeddings live in a feature space compatible enough with the segmentation network's visual features that reducing KL divergence between class centroids and text prototypes is a meaningful training signal that reduces real domain shift.

Editorial extensions

If this is right

  • On the single-source ResNet-50 benchmark spanning five generalization settings, PPAR reports 43.1% average mIoU, exceeding SAN-SAW by 4.6 points, PinMem by 2.1, SHADE by 0.6, and the ISW baseline by 6.7.
  • The same training recipe improves results on lightweight backbones: 37.4% average mIoU on ShuffleNet-V2 and 38.2% on MobileNet-V2, and it adds 0.5-1.3 points on top of vision foundation model baselines such as Rein and Tqdm.
  • In multi-source generalization, training on GTAV+SYNTHIA gives 42.3% average mIoU and training on GTAV+SYNTHIA+IDD gives 56.1%, both above PinMem.
  • Plugging the PPAR modules into SW, IBN, ISW, and SHADE raises their average mIoU by 7.5, 4.1, 5.0, and 1.0 points respectively, indicating the mechanism transfers across methods.
  • Because the CLIP text encoder is used only for prototype generation during training, the deployed segmentation model has no extra inference overhead.

Reading between the lines

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

  • An implicit extension is that frozen text embeddings from a large multimodal model can act as pseudo-target class centroids for other dense prediction tasks with fixed label sets, such as depth estimation or detection; the same OTP/VTP plus reweighting recipe could be tested there.
  • The paper fixes the schedule as shallow-to-VTP and deep-to-OTP; a natural variant it does not explore is selecting the alignment layer per class based on measured shift, or interpolating continuously from VTP to OTP over training.
  • The reweighting term assumes similarity to CLIP text prototypes tracks how transferable a source pixel is; stress-testing on domains far from CLIP's training data, such as fisheye, aerial, or medical imagery, would show whether the gains shrink or reverse.
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 / 6 minor

Summary. The manuscript proposes Prototypical Progressive Alignment and Reweighting (PPAR), a method for single-domain and multi-source domain generalization in semantic segmentation. The method replaces source-computed class prototypes with two fixed text prototypes derived from the CLIP text encoder: an Original Text Prototype (OTP) and a Visual Text Prototype (VTP) that concatenates class names with mode color and LBP texture strings. A prototypical progressive alignment loss aligns shallow features to the VTP and deep features to the OTP using Kullback-Leibler divergences, and a prototypical reweighting loss down-weights pixels whose features are dissimilar to the OTP. The paper also presents a theoretical discussion connecting the method to the G2DM domain-generalization bound. Experiments cover six datasets, three small backbones, multi-source settings, and integration with vision foundation models, reporting an average mIoU of 43.1% on the ResNet-50 single-domain benchmark and further gains on MobileNet-V2, ShuffleNet-V2, and VFM-based backbones.

Significance. If the central claims hold, PPAR would be a useful step for generalizable semantic segmentation: using pre-trained CLIP text prototypes as external, source-independent alignment targets is a plausible way to avoid the known bias of source-computed prototypes, and the easy-to-difficult shallow/deep alignment idea is well motivated by prior evidence that shallow features encode style. The empirical study is broad and internally consistent, with ablations for each component, comparisons against many recent methods, integration into existing baselines, and results on multiple backbones and datasets. The authors also provide a limitation section and do not claim access to target data. However, the paper's central training loss is not fully specified as written, and the reported improvements over several strong baselines are often very small while all tables report single-run mIoU without variance or significance tests. The theoretical contribution is also presented more strongly in the abstract than the actual derivation supports.

major comments (4)
  1. [§IV-E, Eq. (18)] The prototypical progressive alignment loss is not computable as written. Equation (12) computes a KL divergence between the shallow class centroid P_s and the CLIP-based VTP P_tv, and Eq. (13) does the same for the deep centroid and the OTP, but no projection or normalization is described. For ResNet-50, stage-1 (conv2_x) features have 64 channels while CLIP ViT-B/32 text embeddings are 512-dimensional, so the objects in Eq. (12) have different dimensionalities; Table VIII only mentions mapping the last-layer feature dimension to 512, not the stage-1 feature. Moreover, a KL divergence requires non-negative inputs that can be interpreted as distributions (or at least a defined density ratio); raw CNN centroids and raw text embeddings are not such inputs. The same issue affects the dot-product similarity in Eq. (14). Since the ablations in Table VI attribute the largest single-module gains to the progressive alignment loss, the paper must specify the exact projection, normalization, and distributional interpretation used for both P_s and P_d (and for the reweighting similarity), otherwise the method is not reproducible and the stated loss is undefined.
  2. [General experimental reporting] All experimental tables report single-run mIoU without error bars or significance tests, and several of the claimed advantages over state-of-the-art methods are very small. Examples include the 0.5-point margin over PinMem in Table III (42.3 vs. 41.8), the 0.5-point margin over Rein with Dinov2 in Table V (64.8 vs. 64.3), and the 0.6-point margin over DIIA on ShuffleNet-V2 in Table II (37.4 vs. 36.8). For margins of this size, run-to-run variation typical of semantic segmentation training could change the ranking. The claim that PPAR 'surpasses' these methods therefore needs either multiple seeds with means and standard deviations, or a paired significance test, at least for the headline comparisons in Tables I, II, III, and V. Without this, the empirical SOTA claim is not supported at the reported precision.
  3. [Theoretical contribution] The theoretical analysis is presented in the abstract as 'a theoretical analysis showing the alignment between our method and domain generalization theory,' but the actual content is a restatement of the G2DM bound followed by a qualitative argument that reweighting lowers source risk and progressive alignment lowers inter-source divergence. No derivation links the specific losses in Eqs. (12), (13), or (17) to the terms r_i_s or δ of the bound; for instance, it is not shown that the KL losses minimize an inter-domain divergence measure, nor that the reweighted cross-entropy is an unbiased surrogate for source risk. The section is a plausibility discussion rather than a theorem. I recommend either providing a concrete derivation that shows how the proposed losses bound or reduce the terms in Eq. (18), or softening the claim in the abstract and contributions to 'theoretical insight' rather than 'analysis showing alignment.'
  4. [Text prototype generation] The VTP generation procedure is underspecified in ways that affect reproducibility. It is not stated in what color space the mode color C_n is computed, how the mode is taken across spatial locations and channels for a class (e.g., per-channel mode versus joint RGB mode), what the output of LBP is and how its mode is computed, or how the strings resulting from these operations format into the text prompt in Eq. (9). Since the VTP is one of the two central alignment targets, the exact prompt construction and extraction statistics should be described precisely. This is a reproducibility issue for a core component, though it is secondary to the undefined loss in Eqs. (12)-(13).
minor comments (6)
  1. [Implementation details] The paper defines 'Stage 1-4 features' as conv2_x through conv5_x for ResNet, but the same terminology is used for MobileNet-V2 and ShuffleNet-V2 backbones without specifying what 'stage 1' means for those architectures; please clarify how the shallow alignment layer is selected for non-ResNet backbones.
  2. [Typo] There is a typo in the text: 'SAHDE [64]' should be 'SHADE [64]'.
  3. [Figure caption] The caption says 'GT5V' but the dataset is GTA V; please correct to 'GTA V'.
  4. [Reference formatting] Several references contain 'NuerIPS' rather than 'NeurIPS' (e.g., Refs. [18] and [80]); please correct the venue names.
  5. [Hyperparameter] The hyperparameter α_pa is set to 0.001, but no sensitivity analysis or discussion of this choice is provided; a small sensitivity study or a statement that results are stable over a range would strengthen the paper.
  6. [VFM experiments] The description of the VFM experiments is abbreviated: it is not fully clear how 'Ours' is attached to ViT, Dinov2, and Eva-02 backbones, whether the VFM backbones are frozen or fine-tuned, and how 'Rein + Ours' relates to 'Ours' on the Dinov2 backbone; a short experimental clarification would improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: CLIP text prototypes are fixed external anchors, the reweighting uses the same frozen prototypes, and the theoretical bound is a post hoc standard inequality.

full rationale

The derivation chain is self-contained against external anchors rather than circular. Equations (6)-(10) generate OTP and VTP as frozen outputs of the pretrained CLIP text encoder from class names and source color/texture statistics; neither prototype is learned from or chosen to reproduce the reported mIoU. Equations (12)-(13) are auxiliary KL losses pulling class centroids toward those fixed prototypes, and Equations (14)-(17) reweight the cross-entropy loss by similarity to the same frozen OTP; the reweighting values are not fitted to target benchmarks. The theoretical bound in Section IV-E is a standard domain generalization inequality, used post hoc to interpret the losses rather than derived from them, so it cannot be a self-imported uniqueness constraint. Self-citations ([32], [70], [71], [72], [84], [89]) appear only as related work, baseline comparisons, or motivational context and are never the load-bearing justification for the central claim; the strongest reported results are measured against external SOTA methods and public datasets. The skeptical observation that Equations (12)-(13) lack an explicit projection or normalization is a technical reproducibility concern about whether the loss is well-defined as written, not a demonstration that the predictions reduce to the inputs. There is no fitted parameter renamed as a prediction and no equation that is definitionally equal to itself, so no circular step is present.

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

The central contribution rests on one hand-picked hyperparameter (alpha_pa), six domain assumptions about CLIP text space and layer semantics, and one constructed representation (VTP). No new physical entities or fitted constants are introduced.

free parameters (1)
  • alpha_pa = 0.001
    Weight of the prototypical progressive alignment loss in Eq. (5). Chosen by hand, not learned or swept in the paper, and it directly controls the balance between segmentation loss and alignment loss.
assumptions (6)
  • domain assumption CLIP text embeddings define a domain-invariant space whose class centroids are closer to the true overall-domain prototype than source-batch prototypes.
    Section IV.B states CLIP is trained on large-scale wild data and therefore provides high-generalized prototypes. This is the load-bearing premise of both alignment and reweighting.
  • domain assumption The VTP, produced by concatenating mode color and LBP texture strings to class names, lies on an intermediate domain between source and overall target such that shallow alignment is easier than deep alignment.
    Section IV.C and Fig. 3; no quantitative evidence that the constructed text embedding is semantically intermediate beyond intuitive appeal.
  • domain assumption Shallow features encode style and basic visual information while deep features encode category information, so aligning shallow features to VTP and deep features to OTP is appropriate.
    Invoked in Section IV.C from prior works [21], [22]; used to decide which prototype aligns with which layer.
  • standard math The covariance-shift assumption holds and the convex-hull representation of unseen targets applies, reducing the target-risk bound to source risk plus source-domain divergence.
    Section IV.E follows G2DM [86] and Ben-David et al. [85]; standard in the DG literature but an assumption about real urban scene distributions.
  • domain assumption Pixel-level softmax similarity to OTP is a valid proxy for feature reliability and transferability.
    Section IV.D, Eqs. (14)-(17); the entire reweighting mechanism depends on this proxy.
  • domain assumption Visual features and CLIP text features are comparable via KL divergence after a dimension mapping, including for shallow features.
    Eqs. (12)-(13) require distributions over the same support; the paper only notes in Table VIII that the last-layer dimension is mapped to 512, with no description of the projection for Stage 1 features.
invented entities (1)
  • Visual Text Prototype (VTP)
    purpose: Intermediate text prototype formed by injecting mode color and LBP texture strings into class names, used as the easy alignment target for shallow features.
    It is a new representation introduced by this paper. It is reproducible from CLIP and source data but has no falsifiable prediction outside the method itself.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Prototypical Progressive Alignment and Reweighting for Generalizable Semantic Segmentation." pith.science (2026). https://pith.science/paper/LPX6HSLC

@misc{pith2026250711955,
  author       = {Pith},
  title        = {Pith review of: Prototypical Progressive Alignment and Reweighting for Generalizable Semantic Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LPX6HSLC}},
  note         = {Machine review of arXiv:2507.11955}
}
read the original abstract

Generalizable semantic segmentation aims to perform well on unseen target domains, a critical challenge due to real-world applications requiring high generalizability. Class-wise prototypes, representing class centroids, serve as domain-invariant cues that benefit generalization due to their stability and semantic consistency. However, this approach faces three challenges. First, existing methods often adopt coarse prototypical alignment strategies, which may hinder performance. Second, naive prototypes computed by averaging source batch features are prone to overfitting and may be negatively affected by unrelated source data. Third, most methods treat all source samples equally, ignoring the fact that different features have varying adaptation difficulties. To address these limitations, we propose a novel framework for generalizable semantic segmentation: Prototypical Progressive Alignment and Reweighting (PPAR), leveraging the strong generalization ability of the CLIP model. Specifically, we define two prototypes: the Original Text Prototype (OTP) and Visual Text Prototype (VTP), generated via CLIP to serve as a solid base for alignment. We then introduce a progressive alignment strategy that aligns features in an easy-to-difficult manner, reducing domain gaps gradually. Furthermore, we propose a prototypical reweighting mechanism that estimates the reliability of source data and adjusts its contribution, mitigating the effect of irrelevant or harmful features (i.e., reducing negative transfer). We also provide a theoretical analysis showing the alignment between our method and domain generalization theory. Extensive experiments across multiple benchmarks demonstrate that PPAR achieves state-of-the-art performance, validating its effectiveness.

Figures

Figures reproduced from arXiv: 2507.11955 by the authors.

Figure 1
Figure 1. Approach comparison. The red ellipsoid denotes the source domain, [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Flowchart of the proposed PPAR method including text prototype generation, prototypical progressive alignment, and prototypical reweighting. Blue [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Illustration of VTP generation. The color and local texture information [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (5 more)
Figure 5
Figure 5. Figure 5: Illustration of prototypical reweighting. OTP is utilized as class-wise [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Qualitative examples using ResNet-50 backbone on the task of GTAV generalizing to [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Qualitative comparison with other SOTA methods using ResNet-50 backbone on the task of G [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Confusion matrix comparison between the baseline and ours in the [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: T-SNE visualization comparison between the baseline and ours in [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

114 extracted references · 74 canonical work pages

  1. [1]

    Threshold-adaptive unsu- pervised focal loss for domain adaptation of semantic segmentation,

    W. Yan, Y . Qian, C. Wang, and M. Yang, “Threshold-adaptive unsu- pervised focal loss for domain adaptation of semantic segmentation,” IEEE Trans. Intell. Transport. Sys. , vol. 24, no. 1, pp. 752–763, 2022

    2022

  2. [2]

    Sfnet-n: An improved sfnet algorithm for semantic segmentation of low-light autonomous driving road scenes,

    H. Wang, Y . Chen, Y . Cai, L. Chen, Y . Li, M. A. Sotelo, and Z. Li, “Sfnet-n: An improved sfnet algorithm for semantic segmentation of low-light autonomous driving road scenes,” IEEE Trans. Intell. Transport. Sys., vol. 23, no. 11, pp. 21 405–21 417, 2022

    2022

  3. [3]

    Multiple relational learning network for joint referring expression comprehension and segmentation,

    G. Hua, M. Liao, S. Tian, Y . Zhang, and W. Zou, “Multiple relational learning network for joint referring expression comprehension and segmentation,” IEEE Trans. Multimedia , 2023

    2023

  4. [4]

    Contrastive tokens and label acti- vation for remote sensing weakly supervised semantic segmentation,

    Z. Hu, J. Gao, Y . Yuan, and X. Li, “Contrastive tokens and label acti- vation for remote sensing weakly supervised semantic segmentation,” IEEE Transactions on Geoscience and Remote Sensing , 2024

    2024

  5. [5]

    Improving robustness of single image super-resolution models with monte carlo method,

    C. Yang, J. Xiao, Y . Ju, G. Qiu, and K.-M. Lam, “Improving robustness of single image super-resolution models with monte carlo method,” in IEEE Inter. Conf. Image Process. (ICIP). IEEE, 2023, pp. 2135–2139

    2023

  6. [6]

    Token contrast for weakly- supervised semantic segmentation,

    L. Ru, H. Zheng, Y . Zhan, and B. Du, “Token contrast for weakly- supervised semantic segmentation,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , 2023, pp. 3093–3102

    2023

  7. [7]

    Exploring more concentrated and consistent activation regions for cross-domain semantic segmentation,

    M. Liao, G. Hua, S. Tian, Y . Zhang, W. Zou, and X. Li, “Exploring more concentrated and consistent activation regions for cross-domain semantic segmentation,” Neurocomputing, vol. 500, pp. 938–948, 2022

    2022

  8. [8]

    Transfer beyond the field of view: Dense panoramic semantic segmentation via unsupervised domain adaptation,

    J. Zhang, C. Ma, K. Yang, A. Roitberg, K. Peng, and R. Stiefelha- gen, “Transfer beyond the field of view: Dense panoramic semantic segmentation via unsupervised domain adaptation,” IEEE Trans. Intell. Transport. Sys., vol. 23, no. 7, pp. 9478–9491, 2021

    2021

Show all 114 references
  1. [9]

    Dual geometric perception for cross-domain road segmentation,

    W. Zou, R. Long, Y . Zhang, M. Liao, Z. Zhou, and S. Tian, “Dual geometric perception for cross-domain road segmentation,” Displays, vol. 76, p. 102332, 2023

    2023

  2. [10]

    Fda: Fourier domain adaptation for semantic segmentation,

    Y . Yang and S. Soatto, “Fda: Fourier domain adaptation for semantic segmentation,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recog- nit. (CVPR), 2020, pp. 4085–4095

    2020

  3. [11]

    Feature-based style randomization for domain generalization,

    Y . Wang, L. Qi, Y . Shi, and Y . Gao, “Feature-based style randomization for domain generalization,” IEEE Trans. Cir. Sys. Video Tech., vol. 32, no. 8, pp. 5495–5509, 2022

    2022

  4. [12]

    Generalizing to unseen domains: A survey on domain generalization,

    J. Wang, C. Lan, C. Liu, Y . Ouyang, T. Qin, W. Lu, Y . Chen, W. Zeng, and P. Yu, “Generalizing to unseen domains: A survey on domain generalization,” IEEE Trans. Know. Data Eng. , 2022

    2022

  5. [13]

    Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data,

    X. Yue, Y . Zhang, S. Zhao, A. Sangiovanni-Vincentelli, K. Keutzer, and B. Gong, “Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV) , 2019, pp. 2100–2110

    2019

  6. [14]

    Fsdr: Frequency space domain randomization for domain generalization,

    J. Huang, D. Guan, A. Xiao, and S. Lu, “Fsdr: Frequency space domain randomization for domain generalization,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , 2021, pp. 6891–6902

    2021

  7. [15]

    Batch normalization: Accelerating deep network training by reducing internal covariate shift,

    S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep network training by reducing internal covariate shift,” in Proc. Int. Conf. Mach. Learn. (ICML) . pmlr, 2015, pp. 448–456

    2015

  8. [16]

    Switchable whitening for deep representation learning,

    X. Pan, X. Zhan, J. Shi, X. Tang, and P. Luo, “Switchable whitening for deep representation learning,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2019, pp. 1863–1871

    2019

  9. [17]

    Bapa-net: Boundary adaptation and prototype alignment for cross-domain semantic segmen- tation,

    Y . Liu, J. Deng, X. Gao, W. Li, and L. Duan, “Bapa-net: Boundary adaptation and prototype alignment for cross-domain semantic segmen- tation,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV) , 2021, pp. 8801–8811

    2021

  10. [18]

    Category anchor-guided unsupervised domain adaptation for semantic segmentation,

    Q. ZHANG, J. Zhang, W. Liu, and D. Tao, “Category anchor-guided unsupervised domain adaptation for semantic segmentation,” Proc. Adv. Neural Inf. Process. Syst. (NuerIPS) , vol. 32, pp. 435–445, 2019

    2019

  11. [19]

    Proto- typical contrast adaptation for domain adaptive semantic segmentation,

    Z. Jiang, Y . Li, C. Yang, P. Gao, Y . Wang, Y . Tai, and C. Wang, “Proto- typical contrast adaptation for domain adaptive semantic segmentation,” in Proc. Eur. Conf. Comput. Vis. (ECCV). Springer, 2022, pp. 36–54

    2022

  12. [20]

    Bi-directional contrastive learning for domain adaptive semantic segmentation,

    G. Lee, C. Eom, W. Lee, H. Park, and B. Ham, “Bi-directional contrastive learning for domain adaptive semantic segmentation,” in Proc. Eur. Conf. Comput. Vis. (ECCV) . Springer, 2022, pp. 38–55

    2022

  13. [21]

    Image style transfer using convolutional neural networks,

    L. A. Gatys, A. S. Ecker, and M. Bethge, “Image style transfer using convolutional neural networks,” in Proceedings of the IEEE Conf. Comp.Vis. Patt. Recognit., 2016, pp. 2414–2423

    2016

  14. [22]

    Fully convolutional adaptation networks for semantic segmentation,

    Y . Zhang, Z. Qiu, T. Yao, D. Liu, and T. Mei, “Fully convolutional adaptation networks for semantic segmentation,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , 2018, pp. 6810–6818

    2018

  15. [23]

    Contextual-relation consis- tent domain adaptation for semantic segmentation,

    J. Huang, S. Lu, D. Guan, and X. Zhang, “Contextual-relation consis- tent domain adaptation for semantic segmentation,” in Proc. Eur. Conf. Comput. Vis. (ECCV) , 2020, pp. 705–722

    2020

  16. [24]

    Scale variance minimization for unsupervised domain adaptation in image segmentation,

    D. Guan, J. Huang, S. Lu, and A. Xiao, “Scale variance minimization for unsupervised domain adaptation in image segmentation,” Pattern Recognition, vol. 112, p. 107764, 2021

    2021

  17. [25]

    Dast: Unsupervised domain adaptation in semantic segmentation based on discriminator attention and self-training,

    F. Yu, M. Zhang, H. Dong, S. Hu, B. Dong, and L. Zhang, “Dast: Unsupervised domain adaptation in semantic segmentation based on discriminator attention and self-training,” in Proc. AAAI Conf. Artif. Int, vol. 35, no. 12, 2021, pp. 10 754–10 762

    2021

  18. [26]

    Characterizing and avoiding negative transfer,

    Z. Wang, Z. Dai, B. P ´oczos, and J. Carbonell, “Characterizing and avoiding negative transfer,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2019, pp. 11 293–11 302

    2019

  19. [27]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark et al., “Learning transferable visual models from natural language supervision,” in Proc. Int. Conf. Mach. Learn. (ICML) . PMLR, 2021, pp. 8748–8763

    2021

  20. [28]

    Curriculum domain adaptation for semantic segmentation of urban scenes,

    Y . Zhang, P. David, and B. Gong, “Curriculum domain adaptation for semantic segmentation of urban scenes,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV) , 2017, pp. 2020–2030

    2017

  21. [29]

    Map-guided curriculum domain adaptation and uncertainty-aware evaluation for semantic nighttime image segmentation,

    C. Sakaridis, D. Dai, and L. Van Gool, “Map-guided curriculum domain adaptation and uncertainty-aware evaluation for semantic nighttime image segmentation,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 44, no. 6, pp. 3139–3153, 2020

    2020

  22. [30]

    Adversarial domain adaptation with domain mixup,

    M. Xu, J. Zhang, B. Ni, T. Li, C. Wang, Q. Tian, and W. Zhang, “Adversarial domain adaptation with domain mixup,” in Proc. AAAI Conf. Artif. Int , vol. 34, no. 04, 2020, pp. 6502–6509

    2020

  23. [31]

    Dual mixup regularized learning for adversarial domain adaptation,

    Y . Wu, D. Inkpen, and A. El-Roby, “Dual mixup regularized learning for adversarial domain adaptation,” in Proc. Eur. Conf. Comput. Vis. (ECCV), 2020, pp. 540–555

    2020

  24. [32]

    A hybrid domain learning framework for unsupervised semantic segmentation,

    Y . Zhang, S. Tian, M. Liao, W. Zou, and C. Xu, “A hybrid domain learning framework for unsupervised semantic segmentation,” Neuro- computing, vol. 516, pp. 133–145, 2023

    2023

  25. [33]

    Deep dual-resolution networks for real-time and accurate semantic segmentation of traffic scenes,

    H. Pan, Y . Hong, W. Sun, and Y . Jia, “Deep dual-resolution networks for real-time and accurate semantic segmentation of traffic scenes,” IEEE Trans. Intell. Transport. Sys. , vol. 24, no. 3, pp. 3448–3460, 2022. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 15

    2022

  26. [34]

    Delivering arbitrary-modal semantic segmenta- tion,

    J. Zhang, R. Liu, H. Shi, K. Yang, S. Reiß, K. Peng, H. Fu, K. Wang, and R. Stiefelhagen, “Delivering arbitrary-modal semantic segmenta- tion,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2023, pp. 1136–1147

    2023

  27. [35]

    Fully convolutional networks for semantic segmentation,

    J. Long, E. Shelhamer, and T. Darrell, “Fully convolutional networks for semantic segmentation,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2015, pp. 3431–3440

    2015

  28. [36]

    Segnet: A deep convolutional encoder-decoder architecture for image segmentation,

    V . Badrinarayanan, A. Kendall, and R. Cipolla, “Segnet: A deep convolutional encoder-decoder architecture for image segmentation,” IEEE Trans. Pattern Anal. Mach. Intell. , vol. 39, no. 12, pp. 2481– 2495, 2017

    2017

  29. [37]

    Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs,

    L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and A. L. Yuille, “Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs,” IEEE Trans. Pattern Anal. Mach. Intell. , vol. 40, no. 4, pp. 834–848, 2017

    2017

  30. [38]

    Rethinking atrous convolution for semantic image segmentation,

    L.-C. Chen, G. Papandreou, F. Schroff, and H. Adam, “Rethinking atrous convolution for semantic image segmentation,” arXiv preprint arXiv:1706.05587, 2017

    2017 arXiv

  31. [39]

    Encoder-decoder with atrous separable convolution for semantic im- age segmentation,

    L.-C. Chen, Y . Zhu, G. Papandreou, F. Schroff, and H. Adam, “Encoder-decoder with atrous separable convolution for semantic im- age segmentation,” in Proc. Eur. Conf. Comput. Vis. (ECCV), 2018, pp. 801–818

    2018

  32. [40]

    Densely connected convolutional networks,

    G. Huang, Z. Liu, L. Van Der Maaten, and K. Q. Weinberger, “Densely connected convolutional networks,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , 2017, pp. 4700–4708

    2017

  33. [41]

    Deep high-resolution represen- tation learning for human pose estimation,

    K. Sun, B. Xiao, D. Liu, and J. Wang, “Deep high-resolution represen- tation learning for human pose estimation,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , 2019, pp. 5693–5703

    2019

  34. [42]

    Lite-hrnet: A lightweight high-resolution network,

    C. Yu, B. Xiao, C. Gao, L. Yuan, L. Zhang, N. Sang, and J. Wang, “Lite-hrnet: A lightweight high-resolution network,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , 2021, pp. 10 440–10 450

    2021

  35. [43]

    Segnext: Rethinking convolutional attention design for semantic segmentation,

    M.-H. Guo, C.-Z. Lu, Q. Hou, Z.-N. Liu, M.-M. Cheng, and S.-m. Hu, “Segnext: Rethinking convolutional attention design for semantic segmentation,” in Proc. Adv. Neural Inf. Process. Syst. (NuerIPS)

  36. [44]

    Segmenter: Trans- former for semantic segmentation,

    R. Strudel, R. Garcia, I. Laptev, and C. Schmid, “Segmenter: Trans- former for semantic segmentation,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV) , 2021, pp. 7262–7272

    2021

  37. [45]

    Multi-scale high-resolution vision transformer for semantic segmentation,

    J. Gu, H. Kwon, D. Wang, W. Ye, M. Li, Y .-H. Chen, L. Lai, V . Chan- dra, and D. Z. Pan, “Multi-scale high-resolution vision transformer for semantic segmentation,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2022, pp. 12 094–12 103

    2022

  38. [46]

    Gcnet: Non-local networks meet squeeze-excitation networks and beyond,

    Y . Cao, J. Xu, S. Lin, F. Wei, and H. Hu, “Gcnet: Non-local networks meet squeeze-excitation networks and beyond,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV) workshops , 2019, pp. 0–0

    2019

  39. [47]

    Ccnet: Criss-cross attention for semantic segmentation,

    Z. Huang, X. Wang, L. Huang, C. Huang, Y . Wei, and W. Liu, “Ccnet: Criss-cross attention for semantic segmentation,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV) , 2019, pp. 603–612

    2019

  40. [48]

    Pidnet: A real-time semantic segmentation network inspired by pid controllers,

    J. Xu, Z. Xiong, and S. P. Bhattacharyya, “Pidnet: A real-time semantic segmentation network inspired by pid controllers,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2023, pp. 19 529–19 539

    2023

  41. [49]

    Erfnet: Effi- cient residual factorized convnet for real-time semantic segmentation,

    E. Romera, J. M. Alvarez, L. M. Bergasa, and R. Arroyo, “Erfnet: Effi- cient residual factorized convnet for real-time semantic segmentation,” IEEE Trans. Intell. Transport. Sys. , vol. 19, no. 1, pp. 263–272, 2017

    2017

  42. [50]

    Mscfnet: a lightweight network with multi-scale context fusion for real-time semantic segmentation,

    G. Gao, G. Xu, Y . Yu, J. Xie, J. Yang, and D. Yue, “Mscfnet: a lightweight network with multi-scale context fusion for real-time semantic segmentation,” IEEE Trans. Intell. Transport. Sys. , vol. 23, no. 12, pp. 25 489–25 499, 2021

    2021

  43. [51]

    A multi-phase camera-lidar fusion network for 3d semantic segmentation with weak supervision,

    X. Chang, H. Pan, W. Sun, and H. Gao, “A multi-phase camera-lidar fusion network for 3d semantic segmentation with weak supervision,” IEEE Trans. Cir. Sys. Video Tech. , 2023

    2023

  44. [52]

    Rgb-d semantic segmentation and label-oriented voxelgrid fusion for accurate 3d semantic mapping,

    W. Shi, J. Xu, D. Zhu, G. Zhang, X. Wang, J. Li, and X. Zhang, “Rgb-d semantic segmentation and label-oriented voxelgrid fusion for accurate 3d semantic mapping,” IEEE Trans. Cir. Sys. Video Tech., vol. 32, no. 1, pp. 183–197, 2021

    2021

  45. [53]

    Confidence-and-refinement adaptation model for cross-domain seman- tic segmentation,

    X. Zhang, Y . Chen, Z. Shen, Y . Shen, H. Zhang, and Y . Zhang, “Confidence-and-refinement adaptation model for cross-domain seman- tic segmentation,” IEEE Trans. Intell. Transport. Sys. , vol. 23, no. 7, pp. 9529–9542, 2022

    2022

  46. [54]

    Learning texture invariant representation for domain adaptation of semantic segmentation,

    M. Kim and H. Byun, “Learning texture invariant representation for domain adaptation of semantic segmentation,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2020, pp. 12 975–12 984

    2020

  47. [55]

    Affinity space adaptation for semantic segmentation across domains,

    W. Zhou, Y . Wang, J. Chu, J. Yang, X. Bai, and Y . Xu, “Affinity space adaptation for semantic segmentation across domains,” IEEE Trans. Image Process., vol. 30, pp. 2549–2561, 2020

    2020

  48. [56]

    Confidence regularized self-training,

    Y . Zou, Z. Yu, X. Liu, B. Kumar, and J. Wang, “Confidence regularized self-training,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2019, pp. 5982–5991

    2019

  49. [57]

    Rectifying pseudo label learning via uncer- tainty estimation for domain adaptive semantic segmentation,

    Z. Zheng and Y . Yang, “Rectifying pseudo label learning via uncer- tainty estimation for domain adaptive semantic segmentation,” Inter- national Journal of Computer Vision , vol. 129, no. 4, pp. 1106–1120, 2021

    2021

  50. [58]

    Towards robust semantic segmentation of accident scenes via multi- source mixed sampling and meta-learning,

    X. Luo, J. Zhang, K. Yang, A. Roitberg, K. Peng, and R. Stiefelhagen, “Towards robust semantic segmentation of accident scenes via multi- source mixed sampling and meta-learning,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , 2022, pp. 4429–4439

    2022

  51. [59]

    Dacs: Domain adaptation via cross-domain mixed sampling,

    W. Tranheden, V . Olsson, J. Pinto, and L. Svensson, “Dacs: Domain adaptation via cross-domain mixed sampling,” in Proc. IEEE Winter Conf. Appl. Comput. Vis. (WACV) , 2021, pp. 1379–1389

    2021

  52. [60]

    Context-aware mixup for domain adaptive semantic segmentation,

    Q. Zhou, Z. Feng, Q. Gu, J. Pang, G. Cheng, X. Lu, J. Shi, and L. Ma, “Context-aware mixup for domain adaptive semantic segmentation,” IEEE Trans. Cir. Sys. Video Tech. , 2022

    2022

  53. [61]

    Daformer: Improving network architectures and training strategies for domain-adaptive semantic seg- mentation,

    L. Hoyer, D. Dai, and L. Van Gool, “Daformer: Improving network architectures and training strategies for domain-adaptive semantic seg- mentation,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2022, pp. 9924–9935

    2022

  54. [62]

    Domain- invariant information aggregation for domain generalization semantic segmentation,

    M. Liao, S. Tian, Y . Zhang, G. Hua, W. Zou, and X. Li, “Domain- invariant information aggregation for domain generalization semantic segmentation,” Neurocomputing, p. 126273, 2023

    2023

  55. [63]

    Global and local texture randomization for synthetic-to-real semantic segmentation,

    D. Peng, Y . Lei, L. Liu, P. Zhang, and J. Liu, “Global and local texture randomization for synthetic-to-real semantic segmentation,” IEEE Trans. Image Process. , vol. 30, pp. 6594–6608, 2021

    2021

  56. [64]

    Style- hallucinated dual consistency learning: A unified framework for visual domain generalization,

    Y . Zhao, Z. Zhong, N. Zhao, N. Sebe, and G. H. Lee, “Style- hallucinated dual consistency learning: A unified framework for visual domain generalization,” arXiv preprint arXiv:2212.09068 , 2022

    2022 arXiv

  57. [65]

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

    X. Pan, P. Luo, J. Shi, and X. Tang, “Two at once: Enhancing learning and generalization capacities via ibn-net,” in Proc. Eur. Conf. Comput. Vis. (ECCV), 2018, pp. 464–479

    2018

  58. [66]

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

    S. Choi, S. Jung, H. Yun, J. T. Kim, S. Kim, and J. Choo, “Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2021, pp. 11 580–11 590

    2021

  59. [67]

    Semantic-aware domain generalized segmentation,

    D. Peng, Y . Lei, M. Hayat, Y . Guo, and W. Li, “Semantic-aware domain generalized segmentation,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , 2022, pp. 2594–2605

    2022

  60. [68]

    Generalizable model-agnostic se- mantic segmentation via target-specific normalization,

    J. Zhang, L. Qi, Y . Shi, and Y . Gao, “Generalizable model-agnostic se- mantic segmentation via target-specific normalization,” Pattern Recog- nition, vol. 122, p. 108292, 2022

    2022

  61. [69]

    Pin the memory: Learning to generalize semantic segmentation,

    J. Kim, J. Lee, J. Park, D. Min, and K. Sohn, “Pin the memory: Learning to generalize semantic segmentation,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , 2022, pp. 4350–4360

    2022

  62. [70]

    Fine- grained self-supervision for generalizable semantic segmentation,

    Y . Zhang, S. Tian, M. Liao, Z. Zhang, W. Zou, and C. Xu, “Fine- grained self-supervision for generalizable semantic segmentation,” IEEE Trans. Cir. Sys. Video Tech. , 2023

    2023

  63. [71]

    Class-balanced sampling and discriminative stylization for domain generalization se- mantic segmentation,

    M. Liao, S. Tian, B. Wei, Y . Zhang, W. Zou, and X. Li, “Class-balanced sampling and discriminative stylization for domain generalization se- mantic segmentation,” IEEE Transactions on Intelligent Transportation Systems, 2024

    2024

  64. [72]

    Calibration- based multi-prototype contrastive learning for domain generalization semantic segmentation in traffic scenes,

    M. Liao, S. Tian, Y . Zhang, G. Hua, W. Zou, and X. Li, “Calibration- based multi-prototype contrastive learning for domain generalization semantic segmentation in traffic scenes,” IEEE Transactions on Intel- ligent Transportation Systems, 2024

    2024

  65. [73]

    Cris: Clip-driven referring image segmentation,

    Z. Wang, Y . Lu, Q. Li, X. Tao, Y . Guo, M. Gong, and T. Liu, “Cris: Clip-driven referring image segmentation,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , 2022, pp. 11 686–11 695

    2022

  66. [74]

    Referring image segmentation using text supervision,

    F. Liu, Y . Liu, Y . Kong, K. Xu, L. Zhang, B. Yin, G. Hancke, and R. Lau, “Referring image segmentation using text supervision,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV) , 2023, pp. 22 124–22 134

    2023

  67. [75]

    Unsupervised domain adaptation for referring semantic segmentation,

    H. Shi, W. Pan, Z. Zhao, M. Zhang, and F. Wu, “Unsupervised domain adaptation for referring semantic segmentation,” in Proc. ACM Int. Conf. Multi. (ACMMM) , 2023, pp. 5807–5818

    2023

  68. [76]

    A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model,

    M. Xu, Z. Zhang, F. Wei, Y . Lin, Y . Cao, H. Hu, and X. Bai, “A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model,” in Proc. Eur. Conf. Comput. Vis. (ECCV) . Springer, 2022, pp. 736–753

    2022

  69. [77]

    Groupvit: Semantic segmentation emerges from text supervision,

    J. Xu, S. De Mello, S. Liu, W. Byeon, T. Breuel, J. Kautz, and X. Wang, “Groupvit: Semantic segmentation emerges from text supervision,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , 2022, pp. 18 134–18 144

    2022

  70. [78]

    Open-world semantic segmentation via contrasting and clustering vision-language JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 16 embedding,

    Q. Liu, Y . Wen, J. Han, C. Xu, H. Xu, and X. Liang, “Open-world semantic segmentation via contrasting and clustering vision-language JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 16 embedding,” in Proc. Eur. Conf. Comput. Vis. (ECCV). Springer, 2022, pp. 275–292

    2021

  71. [79]

    Decouplenet: Decoupled network for domain adaptive semantic segmentation,

    X. Lai, Z. Tian, X. Xu, Y . Chen, S. Liu, H. Zhao, L. Wang, and J. Jia, “Decouplenet: Decoupled network for domain adaptive semantic segmentation,” in Proc. Eur. Conf. Comput. Vis. (ECCV) . Springer, 2022, pp. 369–387

    2022

  72. [80]

    Subsidiary prototype alignment for universal domain adaptation,

    J. N. Kundu, S. Bhambri, A. R. Kulkarni, H. Sarkar, V . Jampani, and V . B. Radhakrishnan, “Subsidiary prototype alignment for universal domain adaptation,” in Proc. Adv. Neural Inf. Process. Syst. (NuerIPS)

  73. [81]

    Adaptive refining-aggregation-separation framework for unsupervised domain adaptation semantic segmentation,

    Y . Cao, H. Zhang, X. Lu, Y . Chen, Z. Xiao, and Y . Wang, “Adaptive refining-aggregation-separation framework for unsupervised domain adaptation semantic segmentation,” IEEE Trans. Cir. Sys. Video Tech., 2023

    2023

  74. [82]

    Performance evaluation of texture measures with classification based on kullback discrimination of distributions,

    T. Ojala, M. Pietikainen, and D. Harwood, “Performance evaluation of texture measures with classification based on kullback discrimination of distributions,” in Proc. Int. Conf. Patt. Recognit. , vol. 1. IEEE, 1994, pp. 582–585

    1994

  75. [83]

    Pietik ¨ainen, A

    M. Pietik ¨ainen, A. Hadid, G. Zhao, and T. Ahonen, Computer vision using local binary patterns . Springer Science & Business Media, 2011, vol. 40

    2011

  76. [84]

    A global reweighting approach for cross-domain semantic segmentation,

    Y . Zhang, S. Tian, M. Liao, G. Hua, W. Zou, and C. Xu, “A global reweighting approach for cross-domain semantic segmentation,” Signal Processing: Image Communication , p. 117197, 2024

    2024

  77. [85]

    A theory of learning from different domains,

    S. Ben-David, J. Blitzer, K. Crammer, A. Kulesza, F. Pereira, and J. W. Vaughan, “A theory of learning from different domains,”Machine learning, vol. 79, no. 1, pp. 151–175, 2010

    2010

  78. [86]

    Generalizing to unseen domains via distribution matching,

    I. Albuquerque, J. Monteiro, M. Darvishi, T. H. Falk, and I. Mitliagkas, “Generalizing to unseen domains via distribution matching,” arXiv preprint arXiv:1911.00804, 2019

    1911 arXiv

  79. [87]

    Aadg: automatic augmentation for domain generalization on retinal image segmentation,

    J. Lyu, Y . Zhang, Y . Huang, L. Lin, P. Cheng, and X. Tang, “Aadg: automatic augmentation for domain generalization on retinal image segmentation,” IEEE Trans. Medi.l Ima. , vol. 41, no. 12, pp. 3699– 3711, 2022

    2022

  80. [88]

    Learning shape-invariant representation for generalizable semantic segmenta- tion,

    Y . Zhang, S. Tian, M. Liao, G. Hua, W. Zou, and C. Xu, “Learning shape-invariant representation for generalizable semantic segmenta- tion,” IEEE Trans. Image Process. , 2023

    2023

  81. [89]

    Video generalized semantic segmentation via non-salient feature rea- soning and consistency,

    Y . Zhang, Z. Zhang, M. Liao, S. Tian, R. You, W. Zou, and C. Xu, “Video generalized semantic segmentation via non-salient feature rea- soning and consistency,” Knowledge-Based Systems, p. 111584, 2024

    2024

  82. [90]

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

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

    2023

  83. [91]

    Progres- sive random convolutions for single domain generalization,

    S. Choi, D. Das, S. Choi, S. Yang, H. Park, and S. Yun, “Progres- sive random convolutions for single domain generalization,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , 2023, pp. 10 312–10 322

    2023

  84. [92]

    An information-theoretic method to automatic shortcut avoidance and domain generalization for dense prediction tasks,

    W. Chuah, R. Tennakoon, R. Hoseinnezhad, D. Suter, and A. Bab- Hadiashar, “An information-theoretic method to automatic shortcut avoidance and domain generalization for dense prediction tasks,” IEEE Trans. Pattern Anal. Mach. Intell. , 2023

    2023

  85. [93]

    Order-preserving consistency regularization for domain adaptation and generalization,

    M. Jing, X. Zhen, J. Li, and C. G. Snoek, “Order-preserving consistency regularization for domain adaptation and generalization,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV) , 2023, pp. 18 916–18 927

    2023

  86. [94]

    The cityscapes dataset for semantic urban scene understanding,

    M. Cordts, M. Omran, S. Ramos, T. Rehfeld, M. Enzweiler, R. Be- nenson, U. Franke, S. Roth, and B. Schiele, “The cityscapes dataset for semantic urban scene understanding,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , 2016, pp. 3213–3223

    2016

  87. [95]

    Bdd100k: A diverse driving dataset for heterogeneous multitask learning,

    F. Yu, H. Chen, X. Wang, W. Xian, Y . Chen, F. Liu, V . Madhavan, and T. Darrell, “Bdd100k: A diverse driving dataset for heterogeneous multitask learning,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2020, pp. 2636–2645

    2020

  88. [96]

    The mapillary vistas dataset for semantic understanding of street scenes,

    G. Neuhold, T. Ollmann, S. Rota Bulo, and P. Kontschieder, “The mapillary vistas dataset for semantic understanding of street scenes,” in Proceedings of the IEEE international conference on computer vision , 2017, pp. 4990–4999

    2017

  89. [97]

    Idd: A dataset for exploring problems of autonomous navigation in unconstrained environments,

    G. Varma, A. Subramanian, A. Namboodiri, M. Chandraker, and C. Jawahar, “Idd: A dataset for exploring problems of autonomous navigation in unconstrained environments,” in Proc. IEEE Winter Conf. Appl. Comput. Vis. (WACV) . IEEE, 2019, pp. 1743–1751

    2019

  90. [98]

    The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes,

    G. Ros, L. Sellart, J. Materzynska, D. Vazquez, and A. M. Lopez, “The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2016, pp. 3234–3243

    2016

  91. [99]

    Playing for data: Ground truth from computer games,

    S. R. Richter, V . Vineet, S. Roth, and V . Koltun, “Playing for data: Ground truth from computer games,” in Proc. Eur. Conf. Comput. Vis. (ECCV), 2016, pp. 102–118

    2016

  92. [100]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2016, pp. 770–778

    2016

  93. [101]

    Mo- bilenetv2: Inverted residuals and linear bottlenecks,

    M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.-C. Chen, “Mo- bilenetv2: Inverted residuals and linear bottlenecks,” in Proceedings of the IEEE Conf. Comp.Vis. Patt. Recognit. , 2018, pp. 4510–4520

    2018

  94. [102]

    Shufflenet v2: Practical guidelines for efficient cnn architecture design,

    N. Ma, X. Zhang, H.-T. Zheng, and J. Sun, “Shufflenet v2: Practical guidelines for efficient cnn architecture design,” in Proc. Eur. Conf. Comput. Vis. (ECCV) , 2018, pp. 116–131

    2018

  95. [103]

    Dirl: Domain-invariant representation learning for generalizable semantic segmentation,

    Q. Xu, L. Yao, Z. Jiang, G. Jiang, W. Chu, W. Han, W. Zhang, C. Wang, and Y . Tai, “Dirl: Domain-invariant representation learning for generalizable semantic segmentation,” in Proc. AAAI Conf. Artif. Int, vol. 36, no. 3, 2022, pp. 2884–2892

    2022

  96. [104]

    Learning to generalize: Meta-learning for domain generalization,

    D. Li, Y . Yang, Y .-Z. Song, and T. Hospedales, “Learning to generalize: Meta-learning for domain generalization,” in Proc. AAAI Conf. Artif. Int, vol. 32, no. 1, 2018

    2018

  97. [105]

    Stronger fewer & superior: Harnessing vision founda- tion models for domain generalized semantic segmentation,

    Z. Wei, L. Chen, Y . Jin, X. Ma, T. Liu, P. Ling, B. Wang, H. Chen, and J. Zheng, “Stronger fewer & superior: Harnessing vision founda- tion models for domain generalized semantic segmentation,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , 2024, pp. 28 619–28 630

    2024

  98. [106]

    Textual query-driven mask transformer for domain generalized segmentation,

    B. Pak, B. Woo, S. Kim, D.-h. Kim, and H. Kim, “Textual query-driven mask transformer for domain generalized segmentation,” in European Conference on Computer Vision . Springer, 2024, pp. 37–54

    2024

  99. [107]

    Dinov2: Learning robust visual features without supervision,

    M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khali- dov, P. Fernandez, D. Haziza, F. Massa, A. El-Nouby et al., “Dinov2: Learning robust visual features without supervision,” arXiv preprint arXiv:2304.07193, 2023

    2023 arXiv

  100. [108]

    Eva-clip: Improved training techniques for clip at scale,

    Q. Sun, Y . Fang, L. Wu, X. Wang, and Y . Cao, “Eva-clip: Improved training techniques for clip at scale,” arXiv preprint arXiv:2303.15389, 2023

    2023 arXiv

  101. [109]

    Kill two birds with one stone: Domain generalization for semantic segmentation via network pruning,

    Y . Luo, P. Liu, and Y . Yang, “Kill two birds with one stone: Domain generalization for semantic segmentation via network pruning,” Inter- national Journal of Computer Vision , vol. 133, no. 1, pp. 335–352, 2025

    2025

  102. [110]

    Strong but simple: A baseline for domain generalized dense perception by clip-based transfer learning,

    C. H ¨ummer, M. Schwonberg, L. Zhou, H. Cao, A. Knoll, and H. Gottschalk, “Strong but simple: A baseline for domain generalized dense perception by clip-based transfer learning,” in Proceedings of the Asian Conference on Computer Vision , 2024, pp. 4223–4244

    2024

  103. [111]

    Domain adaptive and generaliz- able network architectures and training strategies for semantic image segmentation,

    L. Hoyer, D. Dai, and L. Van Gool, “Domain adaptive and generaliz- able network architectures and training strategies for semantic image segmentation,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 1, pp. 220–235, 2023

    2023

  104. [112]

    Style- hallucinated dual consistency learning: A unified framework for visual domain generalization,

    Y . Zhao, Z. Zhong, N. Zhao, N. Sebe, and G. H. Lee, “Style- hallucinated dual consistency learning: A unified framework for visual domain generalization,” International Journal of Computer Vision, vol. 132, no. 3, pp. 837–853, 2024

    2024

  105. [113]

    Learning content-enhanced mask transformer for domain generalized urban-scene segmentation,

    Q. Bi, S. You, and T. Gevers, “Learning content-enhanced mask transformer for domain generalized urban-scene segmentation,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 38, no. 2, 2024, pp. 819–827

    2024

  106. [114]

    Visualizing data using t-sne

    L. Van der Maaten and G. Hinton, “Visualizing data using t-sne.” Journal of machine learning research , vol. 9, no. 11, 2008

    2008

Pith tools

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