Pith. sign in

REVIEW 4 major objections 6 minor 76 references

Exploring Probabilistic Modeling Beyond Domain Generalization for 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 Semantic segmentation models can be made to generalize to unseen domains by conditioning their features on a latent domain prior learned through probabilistic diffusion modeling, raising DeepLabV3Plus from 40.62 to 49.41 average mIoU and…

desk verdict PDAF's segmentation gains look real, but the 'probabilistic diffusion' claim is an L2 regression in disguise, and the diffusion steps barely matter. read the letter →

arxiv 2507.21367 v1 pith:SRNJ5PQ3 submitted 2025-07-28 cs.CV

I-Hsiang Chen , Hua-En Chang , Wei-Ting Chen , Jenq-Neng Hwang , Sy-Yen Kuo This is my paper
classification cs.CV
keywords domaingeneralizedsemanticsegmentationlatentpriorprobabilisticdiffusionmodelsvariationalinferencefeature-levelalignmenturbanscenepseudo-targetaugmentation
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

PDAF claims that semantic segmentation models can be made to generalize to unseen domains by explicitly modeling a latent domain prior: a small tensor that encodes what is domain-specific about the current input. The paper derives a variational objective in which paired source and pseudo-target images, the latter produced by photometric augmentation, supervise an extractor that learns the prior, a compensation module that uses the prior to rescale and shift features, and a diffusion-based estimator that recovers the prior at inference when no pair is available. If this holds, the framework says that domain generalization does not need to strip style out of features but can instead condition the segmentation network on a probabilistic description of the target's domain shift. The experiments back this with consistent mIoU gains on four unseen urban datasets and on adverse-weather data.

What carries the argument

The load-bearing object is the Latent Domain Prior (LDP), a latent variable $z \in \mathbb{R}^{c' \times h \times w}$ with $c'=4$ that represents domain-specific variation in the feature maps. The variational posterior $q_\varphi(z \mid x_{t'}, x_s)$ is implemented by the LPE, which concatenates source and pseudo-target features and outputs mean and variance; the DCM is an SFT-style affine feature transform, $\tilde\gamma, \tilde\beta = F_\Psi(\tilde z)$ followed by $\tilde h = \tilde\gamma \odot h \oplus \tilde\beta$, that injects the prior into the network's features; and the DPE runs a $T$-step diffusion denoising process conditioned on target features to reconstruct the prior from Gaussian noise at inference. Eq. (3) ties these together by maximizing the likelihood of pseudo-target predictions while minimizing the KL divergence between the posterior and the prior, with the prior constraint loss $L_{\mathrm{prior}} = \|\hat z_0 - \tilde z\|_2$ replacing the KL term in practice.

What would settle it

Run a single-source DGSS experiment where the unseen target shift is not spanned by photometric augmentation, such as a resolution change or a labeling-protocol change, and check whether the mIoU gain over the unmodified backbone disappears. More directly, compare the distribution of LDPs produced by the LPE on source/pseudo-target pairs with those produced by the DPE on real target images across BDD-100K, Mapillary, GTA5, and SYNTHIA; the framework predicts these match closely enough that the DPE-only estimate drives the gain, so a large mismatch on any target would indicate miscalibration.

Watch

Extended reading notes

Core claim

The paper's central claim is that domain-generalized semantic segmentation should be treated as probabilistic inference over a latent variable $z$ (the Latent Domain Prior, LDP) that captures the shift between source and unseen target domains. Its prediction function is $p_{\theta,\phi}(y_t \mid x_t) = \int p_\theta(y_t \mid x_t, z)\,p_\phi(z \mid x_t)\,dz$, and it optimizes the ELBO of this predictive function, instantiated as Eq. (3) using photometric-augmented pseudo-target pairs. At training time, a Latent Prior Extractor (LPE) parameterizes the variational posterior $q_\varphi(z \mid x_{t'}, x_s)$ as a Gaussian via mean and variance outputs; a Domain Compensation Module (DCM) applies the prior as affine scale-and-shift on the features; and a Diffusion Prior Estimator (DPE) learns to denoise a noisy version of the prior conditioned on target features, so that at test time the prior is estimated from the target image alone. The paper reports that this raises DeepLabV3Plus from 40.62 to 49.41 average mIoU on BDD-100K, Mapillary, GTA5, and SYNTHIA when trained on Cityscapes, and reaches 61.08 average mIoU with Mask2Former-Swin-L.

Load-bearing premise

The load-bearing premise is that photometric augmentation of source images produces pseudo-target images whose domain shift is a faithful surrogate for the real unseen target shifts, so a prior learned from those pairs transfers to test domains.

Editorial extensions

If this is right

  • Any existing segmentation encoder-decoder can be wrapped with PDAF: the paper shows consistent gains on DeepLabV3Plus (ResNet-50) and Mask2Former (Swin-T and Swin-L) trained on Cityscapes and on GTA5.
  • The prior constraint loss and the semantic consistency loss each add performance, and the full objective lifts DeepLabV3Plus from 40.62 to 49.41 average mIoU, indicating that conditioning features on the latent domain prior rather than removing domain information is what drives the gain.
  • At inference, the DPE estimates the prior from a single target image without any paired source sample, and increasing the number of denoising steps improves results up to a plateau, with steps 4 to 32 adding only about 0.07 mIoU.
  • PDAF also improves performance on the ACDC adverse-weather subsets (fog, night, rain, snow), suggesting the learned prior captures corruption-like shifts, not just style variation.
  • The framework is reported as state of the art when combined with Mask2Former-Swin-L, reaching 61.08 average mIoU across the four unseen urban datasets.

Reading between the lines

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

  • A natural extension is to broaden the pseudo-target generator beyond photometric augmentation, for example with style hallucination or geometric perturbations, which would test whether the latent prior can represent shift directions that photometric changes cannot produce.
  • Replacing the L2 prior constraint with an actual KL or Wasserstein divergence, or sampling multiple latents per image, would test whether the reported gains come from the variational structure itself or mostly from the affine feature conditioning.
  • The DPE's inference-time denoising cost can be reduced by exploiting accelerated diffusion sampling, since the ablation shows diminishing returns beyond four steps, so a smaller step budget would likely preserve most of the improvement.
  • The strong ACDC results suggest a direct application to adverse-condition segmentation or image restoration, where the latent domain prior could be interpreted as an explicit corruption descriptor that the segmentation head learns to compensate for.
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 paper proposes PDAF, a framework for domain-generalized semantic segmentation that introduces a Latent Domain Prior (LDP) as a conditioning variable. The method has three components: the Latent Prior Extractor (LPE) estimates the LDP from paired source and pseudo-target features, the Domain Compensation Module (DCM) modulates segmentation features with affine transformations derived from the LDP, and the Diffusion Prior Estimator (DPE) estimates the LDP at test time from target features through a denoising process. The training objective combines a task loss, a semantic-consistency loss, and an L2 prior-constraint loss. Experiments on Cityscapes and GTA5 as source domains, with BDD-100K, Mapillary, GTA5/SYNTHIA, and ACDC as unseen targets, report consistent mIoU improvements over strong baselines with DeepLabV3Plus-Res50 and Mask2Former-Swin-T/Swin-L.

Significance. If the reported gains are robust, the paper makes a useful empirical contribution: conditioning an existing segmentation network on an explicitly estimated latent domain prior yields large, consistent improvements across multiple backbones and target domains, including adverse conditions. The paper also includes thorough ablations of the three modules and loss terms. However, the central theoretical claim is not supported by the implemented objective: the ELBO in Eq. (3) is not actually optimized, because the KL term is replaced by an unproven L2 loss in Eq. (10). In addition, key hyperparameters (c' and T) are selected using the unseen target datasets, which biases the reported numbers. The empirical contribution is promising but the manuscript needs substantial revision to align the theory with the implementation and to establish a target-free model-selection protocol.

major comments (4)
  1. [Sec. 3.3, Eq. (10)-(11)] The prior constraint loss Lprior = ||ẑ0 - z̃||² is an L2 distance between the deterministic DPE output and a single reparameterized LPE sample; it is not the KL divergence in Eq. (3). The loss ignores the LPE variance σ, contains no entropy term, and does not constrain z̃ to a standard normal distribution despite the claim in Sec. 3.2. No derivation is given showing that Eq. (10) equals, bounds, or approximates KL(qφ(z'|x_t',x_s)||pϕ(z'|x_t)). Since Eq. (11) is the actual training objective, the paper does not optimize the ELBO in Eqs. (2)-(3), so the central 'probabilistic diffusion' and 'variational inference' claims are not supported by the implementation. Please provide an explicit derivation of the approximation, implement the true reparameterized KL/ELBO, or reframe the contribution as a denoising-based prior estimator without the ELBO framing.
  2. [Sec. 4.3 and Tables 6-7] The hyperparameters c' and T are chosen by maximizing Avg mIoU on the four unseen target datasets (B, M, G, S). Under the single-source DGSS protocol, no target data may be used for model selection. This target-based selection biases the reported gains in Tables 1-3 and makes the ablations in Tables 6-7 a form of test-set tuning. The authors should specify a target-free selection rule (e.g., a held-out source validation set with synthetic shifts) and report the fixed hyperparameter values used to produce the main results.
  3. [Tables 1-7] All results are reported as single numbers with no variance estimates, and baseline numbers are taken from prior papers with possibly different training pipelines. The margins in Table 1 (e.g., DeepLabV3Plus from 40.62 to 49.41 average mIoU) are large, but without error bars it is impossible to judge whether the improvements exceed seed or training variability. Please report mean and standard deviation over at least three runs for the proposed method, and rerun the strongest baselines under the same training setup for a fair comparison.
  4. [Sec. 3.2 and Sec. 4.3] The only pseudo-target generator is photometric augmentation, so the LDP is supervised exclusively by color/illumination shifts. The claim that PDAF generalizes to game-rendered (GTA5), photorealistic simulation (SYNTHIA), and adverse weather (ACDC) relies on the assumption that photometric shifts span the directions of real domain shift. This assumption is not tested. Please add an analysis of the learned LDP under different augmentation types, or explicitly discuss this limitation and its consequences for the generalization claim.
minor comments (6)
  1. [Eq. (3)] The KL term is written as KL(qφ(z'|x_t',x_s)||pϕ(z'|x_t)), but the text and Eq. (8) condition the prior on pseudo-target features h_{θ,t'}; this notational inconsistency should be fixed.
  2. [Eq. (7)] The reverse-diffusion update uses an undefined ε and does not reference the DPE network Fϕ. The authors should state the exact DDIM/DDPM update used, including how Fϕ predicts the noise and how the T denoising steps are applied.
  3. [Sec. 3.2] The sentence 'we constrain the optimal LDP z̃ to follow a standard normal distribution' is not realized in the loss, because Eq. (10) only matches ẑ0 to z̃. Add an explicit KL term to N(0,I) or remove the claim.
  4. [Table 3] The ACDC evaluation protocol is under-specified: please state which Cityscapes training split is used, which ACDC split is evaluated, and whether the same photometric augmentation is applied.
  5. [Table 7 and Sec. 4.3] The caption says 'timestep' and should say 'timesteps.' Also, the improvement from 1 to 32 denoising steps is only 0.6 mIoU (48.88 to 49.48); this small gain should be discussed in the text, as it bears on the contribution of the diffusion process itself.
  6. [Sec. 4.3] The choice of T=4 as an 'optimal trade-off between efficiency and accuracy' appears to be based on target-domain performance (Table 7); this should be reconciled with the target-free selection requirement of Major Comment 2.

Circularity Check

0 steps flagged · score 2.0 of 10

No definitional circularity: the LDP is estimated per test image rather than fitted once, and the reported gains are benchmarked against external methods; the main weakness is an unsupported L2 proxy for the KL term, which is a correctness gap, not a circular reduction.

full rationale

The central derivation is not circular in the prohibited sense. Eqs. (1)-(3) form a standard variational bound with a conditional prior pφ(z|xt); the LPE implements qφ and the DPE implements pφ, and at test time the DPE output ẑ0 is produced from target features plus sampled Gaussian noise, not retrieved from a fitted table or from the LPE's own output. The training supervision Lprior = ||ẑ0 − z̃||² is a distillation-style regression of DPE onto LPE; although this ties DPE to LPE, it does not make a test-set quantity equal to a training input by construction. The reported mIoU numbers are compared against external baselines and the framework is integrated into pretrained, frozen backbones, so the empirical claim has independent content. However, Sec. 3.3 asserts without derivation that the L2 sample loss in Eq. (10) minimizes the KL divergence in Eq. (3); no ELBO, noise-prediction, or variance/normalization term is present, and Table 7 shows the number of diffusion steps has minor effect (48.88 to 49.48), so the 'probabilistic diffusion' framing is not supported by the implemented objective. This is an omitted derivation and correctness concern rather than the kind of circularity where a prediction reduces to a fitted parameter. Hyperparameters c'=4 and T=4 appear selected with knowledge of target-set performance, which is a protocol concern but not definitional circularity. No load-bearing self-citation chain or imported uniqueness theorem appears; the self-citation [6] in related work is motivational only.

Assumptions & free parameters 5 free parameters · 4 assumptions · 1 invented entities

The central claim rests on the assumption that simulated photometric shifts generalize to real domain shifts, plus an ad hoc substitution of an L2 sample loss for the KL divergence in the stated ELBO. Several hyperparameters, including c' and T, are selected using target-domain metrics.

free parameters (5)
  • LDP channel dimension c' = 4
    Chosen as the best value in Table 6, where average mIoU on the four unseen target datasets is used for selection.
  • Diffusion timesteps T = 4
    Chosen in Table 7 based on average mIoU on unseen targets; more steps give slightly higher mIoU, but T=4 is kept for efficiency.
  • Loss coefficients (lambda_task, lambda_sc, lambda_prior) = (0.5, 0.5, 1.0)
    Hand-set; the paper states that hyperparameter analysis is in the supplementary material, which is not provided.
  • Diffusion noise schedule beta_1 to beta_T = 0.1 to 0.99
    Hand-set linear schedule, standard for small T.
  • Pseudo-target augmentation type = Photometric only
    Chosen as the sole augmentation to generate pseudo-targets; no comparison with other augmentations or real target-domain data is provided.
assumptions (4)
  • standard math The variational posterior q_phi(z|x_t, x_s) is a valid approximation to the true posterior, and the ELBO in Eq. (2) is tight enough for optimization.
    Section 3.1 relies on standard variational inference; this is a standard assumption, though the KL term is later replaced by an L2 loss.
  • domain assumption Photometric augmentation of source images produces pseudo-target images whose domain shift approximates the shifts of real unseen target domains.
    Section 3.2 states that paired source and pseudo-target images simulate latent domain shifts; this is load-bearing for the transfer of the LDP.
  • domain assumption The diffusion prior p_phi(z|x_t) can be approximated by the DDIM-style denoiser F_phi conditioned on target features.
    Section 3.2, Eqs. (7) and (8), assumes this parameterized process matches the true prior distribution without a guarantee.
  • ad hoc to paper Minimizing the L2 distance between DPE and LPE samples approximates the KL divergence in Eq. (3).
    Section 3.3, Eq. (10), states that this reduces the distributional discrepancy, but no derivation links the L2 sample loss to the KL term.
invented entities (1)
  • Latent Domain Prior (LDP) z
    purpose: A latent variable intended to capture domain-specific variation and used to condition features through affine transformations in the DCM.
    The LDP has no falsifiable handle outside the proposed framework; its only support is the internal performance gain reported in the paper's experiments. It is a modeling construct, not a verifiable quantity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Exploring Probabilistic Modeling Beyond Domain Generalization for Semantic Segmentation." pith.science (2026). https://pith.science/paper/SRNJ5PQ3

@misc{pith2026250721367,
  author       = {Pith},
  title        = {Pith review of: Exploring Probabilistic Modeling Beyond Domain Generalization for Semantic Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SRNJ5PQ3}},
  note         = {Machine review of arXiv:2507.21367}
}
read the original abstract

Domain Generalized Semantic Segmentation (DGSS) is a critical yet challenging task, as domain shifts in unseen environments can severely compromise model performance. While recent studies enhance feature alignment by projecting features into the source domain, they often neglect intrinsic latent domain priors, leading to suboptimal results. In this paper, we introduce PDAF, a Probabilistic Diffusion Alignment Framework that enhances the generalization of existing segmentation networks through probabilistic diffusion modeling. PDAF introduces a Latent Domain Prior (LDP) to capture domain shifts and uses this prior as a conditioning factor to align both source and unseen target domains. To achieve this, PDAF integrates into a pre-trained segmentation model and utilizes paired source and pseudo-target images to simulate latent domain shifts, enabling LDP modeling. The framework comprises three modules: the Latent Prior Extractor (LPE) predicts the LDP by supervising domain shifts; the Domain Compensation Module (DCM) adjusts feature representations to mitigate domain shifts; and the Diffusion Prior Estimator (DPE) leverages a diffusion process to estimate the LDP without requiring paired samples. This design enables PDAF to iteratively model domain shifts, progressively refining feature representations to enhance generalization under complex target conditions. Extensive experiments validate the effectiveness of PDAF across diverse and challenging urban scenes.

Figures

Figures reproduced from arXiv: 2507.21367 by the authors.

Figure 1
Figure 1. The proposed PDAF enhances existing segmentation backbones (DeepLabV3Plus [ [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overall architecture of PDAF. PDAF augments a pre-trained segmentation network by introducing LDP modeling to enhance its domain generalization. The LPE learns the optimal LDP by modeling cross-domain relationships between source and pseudo-target domains, while DPE is employed to estimate LDP using only target inputs. Finally, the DCM enhances segmentation network with LDP guidance, refining feature alignment and i… view at source ↗
Figure 3
Figure 3. Schematic diagrams of (a) the Latent Prior Extractor, (b) the Domain Compensation Module and (c) the Diffusion Prior Estimator. diffusion modeling to estimate LDP without paired source and target data. Notably, PDAF is compatible with any ex￾isting segmentation model, allowing frozen pre-trained en￾coders Eϑ and decoders Dϑ to discover LDP, and then fine￾tuning the target network Eθ and Dθ for DGSS. The overall arch… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison between DGSS methods trained on GTAV (G) [52] using DeepLabV3Plus [7] with ResNet-50 [21]. The first, second, and third rows show the predictions for Cityscapes (C) [12], BDD-100K (B) [67], and Mapillary (M) [42], respectively. Method Backbone Tr…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

76 extracted references · 67 canonical work pages

  1. [1]

    Style blind domain generalized semantic segmentation via covariance alignment and semantic consis- tence contrastive learning

    Woojin Ahn, Geun Yeong Yang, Hyun Duck Choi, and Myo Taeg Lim. Style blind domain generalized semantic segmentation via covariance alignment and semantic consis- tence contrastive learning. In CVPR, 2024. 2, 3, 6, 7

    work page 2024

  2. [2]

    Metareg: Towards domain generalization using meta- regularization

    Yogesh Balaji, Swami Sankaranarayanan, and Rama Chel- lappa. Metareg: Towards domain generalization using meta- regularization. In NeurIPS, 2018. 2

    work page 2018

  3. [3]

    Lara: Latents and rays for multi-camera bird’s-eye-view semantic segmen- tation

    Florent Bartoccioni, Éloi Zablocki, Andrei Bursuc, Patrick Pérez, Matthieu Cord, and Karteek Alahari. Lara: Latents and rays for multi-camera bird’s-eye-view semantic segmen- tation. In Conference on robot learning, 2023. 2

    work page 2023

  4. [4]

    Collaborating foundation models for domain generalized semantic segmentation

    Yasser Benigmim, Subhankar Roy, Slim Essid, Vicky Kalo- geiton, and Stéphane Lathuilière. Collaborating foundation models for domain generalized semantic segmentation. In CVPR, 2024. 3

    work page 2024

  5. [5]

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

    Qi Bi, Shaodi You, and Theo Gevers. Learning content- enhanced mask transformer for domain generalized urban- scene segmentation. In AAAI, 2024. 3, 7

    work page 2024

  6. [6]

    Unirestore: Unified perceptual and task-oriented image restoration model using diffusion prior

    I Chen, Wei-Ting Chen, Yu-Wei Liu, Yuan-Chun Chiang, Sy-Yen Kuo, Ming-Hsuan Yang, et al. Unirestore: Unified perceptual and task-oriented image restoration model using diffusion prior. In CVPR, 2025. 3

    work page 2025

  7. [7]

    Rethinking atrous convolution for seman- tic image segmentation

    Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for seman- tic image segmentation. arXiv preprint arXiv:1706.05587 ,

    arXiv

  8. [8]

    Rvsl: Robust vehicle similarity learning in real hazy scenes based on semi-supervised learning

    Wei-Ting Chen, I-Hsiang Chen, Chih-Yuan Yeh, Hao- Hsiang Yang, Hua-En Chang, Jian-Jiun Ding, and Sy-Yen Kuo. Rvsl: Robust vehicle similarity learning in real hazy scenes based on semi-supervised learning. In ECCV, 2022. 2

    work page 2022

Show all 76 references
  1. [9]

    Sjdl-vehicle: Semi-supervised joint defogging learning for foggy vehicle re-identification

    Wei-Ting Chen, I-Hsiang Chen, Chih-Yuan Yeh, Hao- Hsiang Yang, Jian-Jiun Ding, and Sy-Yen Kuo. Sjdl-vehicle: Semi-supervised joint defogging learning for foggy vehicle re-identification. In AAAI, 2022. 2

    2022

  2. [10]

    Schwing, Alexan- der Kirillov, and Rohit Girdhar

    Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexan- der Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR,

  3. [11]

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

    Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via in- stance selective whitening. In CVPR, 2021. 3, 6, 7, 8

    2021

  4. [12]

    The cityscapes dataset for semantic urban scene understanding

    Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR,

  5. [13]

    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. In CVPR, 2009. 2

    2009

  6. [14]

    Diffusion models beat gans on image synthesis

    Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 3

    2021

  7. [15]

    Hgformer: Hierarchical grouping transformer for domain generalized semantic segmentation

    Jian Ding, Nan Xue, Guisong Xia, Bernt Schiele, and Dengxin Dai. Hgformer: Hierarchical grouping transformer for domain generalized semantic segmentation. In CVPR,

  8. [16]

    Domain generalization via model-agnostic learning of semantic features

    Qi Dou, Daniel Coelho de Castro, Konstantinos Kamnitsas, and Ben Glocker. Domain generalization via model-agnostic learning of semantic features. In NeurIPS, 2019. 2

    2019

  9. [17]

    Mark Everingham, S. M. Ali Eslami, Luc Van Gool, Christo- pher K. I. Williams, John M. Winn, and Andrew Zisserman. The pascal visual object classes challenge: A retrospective. IJCV, 2014. 6

    2014

  10. [18]

    Unsupervised domain adaptation by backpropagation

    Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In ICML, 2015. 2

    2015

  11. [19]

    Kleijn, Mengjie Zhang, and David Balduzzi

    Muhammad Ghifary, W. Kleijn, Mengjie Zhang, and David Balduzzi. Domain generalization for object recognition with multi-task autoencoders. In ICCV, 2015. 2

    2015

  12. [20]

    Prompting diffusion representations for cross-domain semantic segmentation

    Rui Gong, Martin Danelljan, Han Sun, Julio Delgado Man- gas, and Luc Van Gool. Prompting diffusion representations for cross-domain semantic segmentation. arXiv preprint arXiv:2307.02138, 2023. 3

    2023 arXiv

  13. [21]

    Zhang, Shaoqing Ren, and Jian Sun

    Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2015. 6, 7, 8

    2015

  14. [22]

    Jonathan Ho, Ajay Jain, and P. Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3

    2020

  15. [23]

    Planning-oriented autonomous driving

    Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In CVPR, 2023. 2

    2023

  16. [24]

    Fsdr: Frequency space domain randomization for domain generalization

    Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Fsdr: Frequency space domain randomization for domain generalization. In CVPR, 2021. 2

    2021

  17. [25]

    Itera- tive normalization: Beyond standardization towards efficient whitening

    Lei Huang, Yi Zhou, Fan Zhu, Li Liu, and Ling Shao. Itera- tive normalization: Beyond standardization towards efficient whitening. In CVPR, 2019. 7, 8

    2019

  18. [26]

    Style projected clustering for domain generalized semantic seg- mentation

    Wei Huang, Chang Wen Chen, Yong Li, Jiacheng Li, Cheng Li, Fenglong Song, Youliang Yan, and Zhiwei Xiong. Style projected clustering for domain generalized semantic seg- mentation. In CVPR, 2023. 2, 3, 7

    2023

  19. [27]

    Batch normalization: Accelerating deep network training by reducing internal co- variate shift

    Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal co- variate shift. In ICML, 2015. 2

    2015

  20. [28]

    Diffusion features to bridge domain gap for semantic segmentation

    Yuxiang Ji, Boyong He, Chenyuan Qu, Zhuoyue Tan, Chuan Qin, and Liaoni Wu. Diffusion features to bridge domain gap for semantic segmentation. arXiv preprint arXiv:2406.00777, 2024. 3

    2024 arXiv

  21. [29]

    Dgin- style: Domain-generalizable semantic segmentation with image diffusion models and stylized semantic control

    Yuru Jia, Lukas Hoyer, Shengyu Huang, Tianfu Wang, Luc Van Gool, Konrad Schindler, and Anton Obukhov. Dgin- style: Domain-generalizable semantic segmentation with image diffusion models and stylized semantic control. In ECCV, 2024. 2, 3

    2024

  22. [30]

    Scedit: Efficient and controllable image diffusion generation via skip connection editing

    Zeyinzi Jiang, Chaojie Mao, Yulin Pan, Zhen Han, and Jingfeng Zhang. Scedit: Efficient and controllable image diffusion generation via skip connection editing. In CVPR,

  23. [31]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2014. 6

    2014

  24. [32]

    Kingma and Max Welling

    Diederik P. Kingma and Max Welling. Auto-encoding vari- ational bayes. In ICLR, 2013. 5

    2013

  25. [33]

    Kingma, Tim Salimans, Ben Poole, and Jonathan Ho

    Diederik P. Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. arXiv preprint arXiv:2107.00630, 2021. 3

    2021 arXiv

  26. [34]

    Wildnet: Learning domain generalized semantic seg- mentation from the wild

    Suhyeon Lee, Hongje Seong, Seongwon Lee, and Euntai Kim. Wildnet: Learning domain generalized semantic seg- mentation from the wild. In CVPR, 2022. 2, 7

    2022

  27. [35]

    Hospedales

    Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M. Hospedales. Learning to generalize: Meta-learning for do- main generalization. In AAAI, 2017. 2

    2017

  28. [36]

    Domain generalization with ad- versarial feature learning

    Haoliang Li, Sinno Jialin Pan, Shiqi Wang, and Alex Chichung Kot. Domain generalization with ad- versarial feature learning. In CVPR, 2018. 2

    2018

  29. [37]

    Deep domain generaliza- tion via conditional invariant adversarial networks

    Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generaliza- tion via conditional invariant adversarial networks. InECCV,

  30. [38]

    Cdformer:when degradation prediction embraces diffusion model for blind image super-resolution

    Qingguo Liu, Chenyi Zhuang, Pan Gao, and Jie Qin. Cdformer:when degradation prediction embraces diffusion model for blind image super-resolution. In CVPR, 2024. 3, 5

    2024

  31. [39]

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

    Yajing Liu, Shijun Zhou, Xiyao Liu, Chunhui Hao, Baojie Fan, and Jiandong Tian. Unbiased faster r-cnn for single- source domain generalized object detection. In CVPR, 2024. 2

    2024

  32. [40]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 6

    2021

  33. [41]

    Adjeroh, and Gi- anfranco Doretto

    Saeid Motiian, Marco Piccirilli, Donald A. Adjeroh, and Gi- anfranco Doretto. Unified deep supervised domain adapta- tion and generalization. In ICCV, 2017. 2

    2017

  34. [42]

    The mapillary vistas dataset for semantic understanding of street scenes

    Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulò, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In ICCV, 2017. 6, 7, 8

    2017

  35. [43]

    Embodied visual active learning for semantic segmentation

    David Nilsson, Aleksis Pirinen, Erik Gärtner, and Cristian Sminchisescu. Embodied visual active learning for semantic segmentation. In AAAI, 2021. 2

    2021

  36. [44]

    Au- tonomous mobile robot navigation independent of road boundary using driving recommendation map

    Yuya Onozuka, Ryosuke Matsumi, and Motoki Shino. Au- tonomous mobile robot navigation independent of road boundary using driving recommendation map. In Int. Conf. on Intel. Robots and Systems, 2021. 2

    2021

  37. [45]

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

    Xingang Pan, Ping Luo, Jianping Shi, and Xiaoou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In ECCV, 2018. 2, 6, 7, 8

    2018

  38. [46]

    Switchable whitening for deep representation learning

    Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaoou Tang, and Ping Luo. Switchable whitening for deep representation learning. In ICCV, 2019. 3, 6, 7, 8

    2019

  39. [47]

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

    Duo Peng, Yinjie Lei, Lingqiao Liu, Pingping Zhang, and Jun Liua. Global and local texture randomization for synthetic-to-real semantic segmentation. IEEE TIP , 2021. 2, 7

    2021

  40. [48]

    Semantic-aware domain generalized segmentation

    Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In CVPR, 2022. 3, 7

    2022

  41. [49]

    Semantic-aware domain generalized segmentation

    Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In CVPR, 2022. 6

    2022

  42. [50]

    Richter, Christo- pher J

    Pablo Pernias, Dominic Rampas, Mats L. Richter, Christo- pher J. Pal, and Marc Aubreville. Wuerstchen: An efficient architecture for large-scale text-to-image diffusion models. arXiv preprint arXiv:2306.00637v2, 2023. 3, 5

    2023 arXiv

  43. [51]

    Lead: Learn- ing decomposition for source-free universal domain adapta- tion

    Sanqing Qu, Tianpei Zou, Lianghua He, Florian Röhrbein, Alois Knoll, Guang Chen, and Changjun Jiang. Lead: Learn- ing decomposition for source-free universal domain adapta- tion. In CVPR, 2024. 2

    2024

  44. [52]

    Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun

    Stephan R. Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In ECCV, 2016. 6, 7, 8

    2016

  45. [53]

    Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer

    Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2021. 3, 5

    2021

  46. [54]

    Germán Ros, Laura Sellart, Joanna Materzynska, David Vázquez, and Antonio M. López. The synthia dataset: A large collection of synthetic images for semantic segmenta- tion of urban scenes. In CVPR, 2016. 6, 7, 8

    2016

  47. [55]

    Acdc: The adverse conditions dataset with correspondences for se- mantic driving scene understanding

    Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Acdc: The adverse conditions dataset with correspondences for se- mantic driving scene understanding. In CVPR, 2021. 7, 8

    2021

  48. [56]

    Learning to optimize domain specific normalization for domain generalization

    Seonguk Seo, Yumin Suh, Dongwan Kim, Jongwoo Han, and Bohyung Han. Learning to optimize domain specific normalization for domain generalization. In ECCV, 2019. 2

    2019

  49. [57]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5

    2010 arXiv

  50. [58]

    Your classifier can secretly suffice multi-source domain adaptation

    Naveen Venkat, Jogendra Nath Kundu, Durgesh Singh, Am- bareesh Revanur, et al. Your classifier can secretly suffice multi-source domain adaptation. NIPS, 2020. 2

    2020

  51. [59]

    Exploiting diffusion prior for real-world image super-resolution

    Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. IJCV, 2024. 3

    2024

  52. [60]

    Recovering realistic texture in image super-resolution by deep spatial feature transform

    Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Recovering realistic texture in image super-resolution by deep spatial feature transform. In CVPR, 2018. 5

    2018

  53. [61]

    Domain generalization guided by large-scale pre-trained priors

    Zongbin Wang, Bin Pan, Shiyu Shen, Tianyang Shi, and Zhenwei Shi. Domain generalization guided by large-scale pre-trained priors. arXiv preprint arXiv:2406.05628, 2024. 2

    2024 arXiv

  54. [62]

    Stronger fewer & superior: Harnessing vision foundation models for domain generalized semantic segmentation

    Zhixiang Wei, Lin Chen, Yi Jin, Xiaoxiao Ma, Tianle Liu, Pengyang Ling, Ben Wang, Huaian Chen, and Jinjin Zheng. Stronger fewer & superior: Harnessing vision foundation models for domain generalized semantic segmentation. In CVPR, 2024. 3

    2024

  55. [63]

    Datasetdm: Synthesizing data with perception anno- tations using diffusion models

    Weijia Wu, Yuzhong Zhao, Hao Chen, Yuchao Gu, Rui Zhao, Yefei He, Hong Zhou, Mike Zheng Shou, and Chunhua Shen. Datasetdm: Synthesizing data with perception anno- tations using diffusion models. NeurIPS, 2023. 3

    2023

  56. [64]

    Diffir: Efficient diffusion model for image restoration

    Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xing- long Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. Diffir: Efficient diffusion model for image restoration. In ICCV, 2023. 3

    2023

  57. [65]

    Dirl: Domain-invariant representation learning for gen- eralizable semantic segmentation

    Qi Xu, Lili Yao, Zhengkai Jiang, Guannan Jiang, Wenqing Chu, Wenhui Han, Wei Zhang, Chengjie Wang, and Ying Tai. Dirl: Domain-invariant representation learning for gen- eralizable semantic segmentation. In AAAI, 2022. 3, 7

    2022

  58. [66]

    Generalized seman- tic segmentation by self-supervised source domain projec- tion and multi-level contrastive learning

    Liwei Yang, Xiang Gu, and Jian Sun. Generalized seman- tic segmentation by self-supervised source domain projec- tion and multi-level contrastive learning. In AAAI, 2023. 2, 3, 7

    2023

  59. [67]

    Bdd100k: A diverse driving dataset for heterogeneous multitask learning

    Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Dar- rell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2018. 1, 6, 7, 8

    2018

  60. [68]

    Sangiovanni-Vincentelli, Kurt Keutzer, and Boqing Gong

    Xiangyu Yue, Yang Zhang, Sicheng Zhao, Alberto L. Sangiovanni-Vincentelli, Kurt Keutzer, and Boqing Gong. Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In ICCV, 2019. 2, 7

    2019

  61. [69]

    Improving the generalization of segmentation foundation model under distribution shift via weakly supervised adaptation

    Haojie Zhang, Yongyi Su, Xun Xu, and Kui Jia. Improving the generalization of segmentation foundation model under distribution shift via weakly supervised adaptation. InCVPR,

  62. [70]

    C3net: Compound conditioned controlnet for multi- modal content generation

    Juntao Zhang, Yuehuai Liu, Yu-Wing Tai, and Chi-Keung Tang. C3net: Compound conditioned controlnet for multi- modal content generation. In CVPR, 2024. 3

    2024

  63. [71]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 3

    2023

  64. [72]

    Mamba as a bridge: Where vision foundation models meet vision language models for domain-generalized semantic segmentation

    Xin Zhang and Robby T Tan. Mamba as a bridge: Where vision foundation models meet vision language models for domain-generalized semantic segmentation. In CVPR, 2025. 3

    2025

  65. [73]

    Fishertune: Fisher- guided robust tuning of vision foundation models for domain generalized segmentation

    Dong Zhao, Jinlong Li, Shuang Wang, Mengyao Wu, Qi Zang, Nicu Sebe, and Zhun Zhong. Fishertune: Fisher- guided robust tuning of vision foundation models for domain generalized segmentation. In CVPR, 2025. 3

    2025

  66. [74]

    Uni-controlnet: All-in-one control to text-to-image diffusion models

    Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models. NeurIPS, 2023. 3

    2023

  67. [75]

    Sebe, and Gim Hee Lee

    Yuyang Zhao, Zhun Zhong, Na Zhao, N. Sebe, and Gim Hee Lee. Style-hallucinated dual consistency learning for domain generalized semantic segmentation. In ECCV, 2022. 2, 3, 7

    2022

  68. [76]

    Zhun Zhong, Yuyang Zhao, Gim Hee Lee, and N. Sebe. Ad- versarial style augmentation for domain generalized urban- scene segmentation. In NeurIPS, 2022. 2, 7

    2022

Pith tools

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