Pith. sign in

REVIEW 5 major objections 7 minor 34 references

Domain Generalized Stereo Matching with Uncertainty-guided Data Augmentation

T0 review · 5 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read This paper claims that training stereo networks on synthetically perturbed per-channel RGB statistics, together with a feature consistency loss, makes them generalize to unseen real-world domains without any target-domain data.

desk verdict Training-only augmentation that likely helps stereo DG, but the uncertainty mechanism needs a color-jitter control before we can say it does the work. read the letter →

arxiv 2508.01303 v1 pith:CAWOA4YC submitted 2025-08-02 cs.CV

classification cs.CV
keywords domaingeneralizationstereomatchingdataaugmentationuncertaintymodelingfeatureconsistencyRGBstatisticssynthetic-to-real
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 tries to establish that a simple, training-only data augmentation can close much of the synthetic-to-real domain gap in stereo matching. It argues that per-image RGB channel means and standard deviations carry domain style, and that perturbing these statistics with Gaussian noise whose variance comes from the training batch generates plausible unseen-domain samples. A feature consistency loss then forces the network to keep matching features stable under such style shifts. If correct, any existing stereo matcher can gain cross-domain robustness without changing its architecture or seeing target data.

What carries the argument

The load-bearing object is the uncertainty-guided augmentation in RGB space: per-channel mean $\mu_c(x)$ and standard deviation $\sigma_c(x)$ are computed per image, their batch-level variances $\sigma^2_{\mu_c}(x)$ and $\sigma^2_{\sigma_c}(x)$ define Gaussian perturbation models $N(0,\sigma^2_{\mu_c}(x))$ and $N(0,\sigma^2_{\sigma_c}(x))$, and sampled perturbations are applied via standardization and re-stylization. The second mechanism is the feature consistency loss $L_{\mathrm{cons}} = \|f_{\mathrm{feat}}(x_L)-f_{\mathrm{feat}}(x^*_L)\|_2 + \|f_{\mathrm{feat}}(x_R)-f_{\mathrm{feat}}(x^*_R)\|_2$, which ties original and augmented features together. Together they force the feature extractor to be insensitive to style while preserving geometry.

What would settle it

Replace the Gaussian perturbation with uniform random color jitter whose range matches the batch variation and retrain the same baselines on SceneFlow; if D1 errors on KITTI, Middlebury, and ETH3D do not rise, then the Gaussian calibration is not what drives the gains. A second check is to measure whether target-domain per-channel statistics actually fall inside the augmentation's sampled range; if many targets fall outside, the claimed coverage of unseen domains is disproved.

Watch

Extended reading notes

Core claim

UgDA-stereo's central claim is that the gap between synthetic training and real testing can be reduced by treating domain shift as uncertainty in low-order color statistics. For each channel, the method computes per-image mean and standard deviation, measures their variation across a batch, and samples Gaussian perturbations around those statistics to produce augmented left-right pairs. Training on these pairs with a consistency loss that penalizes feature differences between original and augmented views yields features that preserve scene structure rather than style. The paper reports that this recipe lowers D1 error across KITTI 2012, KITTI 2015, Middlebury, and ETH3D for PSMNet, GwcNet, and CFNet, and improves robustness on rainy and foggy DrivingStereo splits.

Load-bearing premise

The load-bearing premise is that the per-channel color mean and spread inside one training batch of four images define a Gaussian spread of styles broad enough to include unseen target domains; if the batch statistics are not representative, the augmentation becomes ordinary color jitter.

Editorial extensions

If this is right

  • Stereo networks trained only on SceneFlow can become more reliable on KITTI, Middlebury, and ETH3D without any target-domain data or architectural change.
  • The augmentation is training-only, so deployed models pay no extra inference cost for the improved generalization.
  • The feature consistency loss makes learned representations shift-resistant to style, which should reduce shortcut reliance in other geometry-sensitive dense prediction tasks.
  • Because left and right images receive the same perturbation, epipolar geometry and disparity labels remain valid, so the augmentation can be added to any stereo loss pipeline.
  • Reported DrivingStereo results suggest robustness to weather-induced appearance shift, not only to dataset shift.

Reading between the lines

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

  • Editorial inference: the same Gaussian-calibrated perturbation could be applied to other color- or texture-sensitive regression tasks, such as optical flow or monocular depth, wherever batch statistics are informative about style variation.
  • Editorial inference: a direct test of the mechanism would be to compare against plain per-channel color jitter matched to the same perturbation range; if the Gaussian calibration is doing the work, it should beat jitter when batch statistics differ sharply across domains.
  • Editorial inference: the batch-size dependence is a hidden lever. With batch size 4 the variance estimate is noisy; the paper's reported gains might be larger or unstable at different batch sizes, and a systematic study would clarify the method's operating range.
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

5 major / 7 minor

Summary. This paper proposes UgDA-Stereo, a training-only, architecture-agnostic data-augmentation strategy for domain-generalized stereo matching. The method (i) perturbs the per-channel RGB mean and standard deviation of each training image by Gaussian noise whose variances are estimated from the inter-image statistics of the current batch (Eqs. (3)-(10)), aiming to synthesize style variations of unseen target domains, and (ii) adds a feature-consistency loss (Eq. (11)) that encourages the shared feature extractor to produce similar features for original and augmented views of the same scene. Training on SceneFlow, the authors evaluate the method on KITTI 2012, KITTI 2015, Middlebury, and ETH3D with PSMNet, GwcNet, and CFNet backbones, and report consistent D1/EPE improvements over the baselines and several prior domain-generalization methods (Table 1), a two-component ablation (Table 2), and a robustness study on DrivingStereo under different weather conditions (Table 3).

Significance. If the causal mechanism claimed is correct, this is a useful practical contribution: the module is plug-and-play, has negligible overhead, introduces only one hyperparameter (lambda), and the paper delivers broad empirical coverage (three backbones, four benchmarks, and weather-specific robustness tests) with qualitative feature visualizations. The two-component ablation in Table 2 and the consistency between the ablated values and Table 1's full-model numbers are in the paper's favor, as is the fact that the method requires no target-domain access. Its main limitation is that the uncertainty mechanism has not been isolated from generic photometric perturbation (no fixed-variance or color-jitter control), the perturbation distribution in Eq. (8) is left partially unspecified by the missing positivity constraint, and several highlighted numbers in the DrivingStereo section disagree with Table 3. Because the method is simple and the missing evidence is obtainable with modest experiments, the contribution is promising but the central attribution claim needs verification before publication.

major comments (5)
  1. [Ablation Studies, Table 2] Table 2 isolates only three configurations (no augmentation; UgDA without L_cons; UgDA with L_cons), so it never tests whether the batch-calibrated Gaussian in Eqs. (5)-(8) is what produces the improvement. Any augmentation that randomly shifts per-channel RGB mean and standard deviation expands the source domain and could confer a similar regularization benefit, and with the batch size of 4 reported in the Experiment section, the variances in Eqs. (5)-(6) are very noisy single estimates; no control with fixed-magnitude perturbation of matched size, no standard color-jitter or brightness/contrast baseline, and no sensitivity analysis with respect to batch size is reported. Therefore the central claim that uncertainty-guided perturbation (rather than generic photometric augmentation) drives the D1 reductions in Table 2 is not yet supported.
  2. [Eq. (8)] Eq. (8) defines the perturbed standard deviation as sigma'_c(x) = sigma_c(x) + epsilon_sigma_c * sigma_sigma_c(x) with epsilon_sigma_c ~ N(0,1) and imposes no positivity constraint; when epsilon_sigma_c < -sigma_c(x)/sigma_sigma_c(x), sigma'_c(x) becomes negative and Eq. (9) inverts the contrast of that channel, which is a qualitatively different transformation from the intended style shift. The paper neither clamps sigma'_c(x), nor reports how frequently negative values occur, nor states what the implemented code does in that case; this leaves the actual training distribution only partially specified and blocks exact reproduction. Either a positivity-preserving reparameterization or an explicit clamp with its incidence should be provided.
  3. [Robustness to Anomalous Scenarios] The text in 'Robustness to Anomalous Scenarios' reports UgDA-PSMNet D1 errors of 4.2% (Sunny), 3.1% (Cloudy), 8.6% (Rainy), and 5.3% (Foggy), which disagree with Table 3's values of 4.2, 3.3, 6.5, and 5.7, and the text values do not average to the table's 4.9 mean. The same paragraph's claim that UgDA-PSMNet 'significantly outperforms all baseline methods' in these conditions is also contradicted by FT-PSMNet on Sunny (4.0 vs 4.2) and Cloudy (2.9 vs 3.3), and the fine-tuning protocol for FT-PSMNet and FT-GwcNet is nowhere described. The numbers and the claim should be reconciled and the fine-tuning setup specified.
  4. [Reproducibility Checklist] Several reproducibility-checklist answers are not supported by the manuscript: items 4.2 (range of hyperparameter values tried), 4.4 and 4.5 (code appendix and code release statement), 4.7 (seed setting), 4.10 (number of runs per result), 4.11 (variation or confidence measures), and 4.12 (statistical significance tests) are all marked 'yes', yet the paper contains no hyperparameter search, no code appendix or repository statement, no seed description, no indication of repeated runs, and no error bars or statistical tests. Since Tables 1-3 report single scalar values, the significance claims and the checklist assertions should either be substantiated or corrected.
  5. [Related Work and Table 1] HVT (Chang et al. 2023), the input-level hierarchical augmentation method most closely related to UgDA, is discussed in Related Work but never appears in Table 1 or any other comparison; the same holds for other augmentation-based domain-generalization methods for stereo matching. Because the contribution of this paper is itself an input-level augmentation scheme, the claim that UgDA 'significantly improves the generalization performance of existing SM networks' needs a head-to-head comparison with HVT to be convincing; I recommend adding it to Table 1.
minor comments (7)
  1. [Abstract] The abstract contains typos ('poposed', 'unceratinty', 'Specifcally'), and 'Specifcally' recurs in the Method section; these should be corrected.
  2. [Eq. (4)] Eq. (4) is missing a closing parenthesis in the squared term; it should read (x_{c,h,w} - mc(x))^2.
  3. [Structural Advantages] The claim that the augmentation 'explicitly maintains left-right consistency' needs clarification: since Eqs. (7)-(8) sample independent perturbations for each image, A(x_L) and A(x_R) generally receive different channel-wise transforms, so the paper should state whether the left/right draws are coupled and, if not, why the photometric inconsistency between views is acceptable for stereo matching.
  4. [Learning Domain-Invariant Features] The claim that L_cons keeps feature representations consistent before and after perturbation is supported only by qualitative histograms (Figure 5); a quantitative feature-distance measure would make the claim testable.
  5. [Table 2] The column header of Table 2 ('KITTI-2012 KITTI-2015' followed by 'PSMNet GwcNet PSMNet GwcNet') is ambiguous; the four result columns should be labeled explicitly.
  6. [Figure 4] In Figure 4, the axes, arrows, and their relation to the perturbation direction and amplitude in Eqs. (7)-(8) are not labeled; the caption should define what each arrow represents.
  7. [Experiment] lambda = 0.17 is the only hyperparameter of the method, but no sensitivity analysis is reported; a small sweep would clarify how much of Table 2's gain from L_cons depends on this specific value.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: UgDA-Stereo is a training-time augmentation whose parameters come from the source batch, and all claims are evaluated on unseen target benchmarks.

full rationale

UgDA-Stereo's derivation chain is self-contained: the augmentation parameters in Eqs. (5)-(6) are computed from the source training batch and define a sampling distribution for perturbing per-channel RGB statistics; Eq. (9) is an explicit affine remapping of the input. No target-domain statistic, fitted parameter, or learned quantity is later renamed as a prediction. The improvement claim is an empirical generalization result measured on KITTI, Middlebury, ETH3D, and DrivingStereo, which are outside the SceneFlow training domain. The one cited methodological inspiration for Gaussian uncertainty modeling (Li et al. 2022) is not authored by the present authors and is not used to force a uniqueness result; the paper's own equations are what define the method. The lack of an ablation replacing the batch-calibrated Gaussian with fixed-variance color jitter means the specific role of 'uncertainty guidance' is under-supported experimentally, but that is an attribution and experimental-design concern rather than a circularity: the paper never defines the evaluated metric in terms of the augmentation's own statistics. Reproducibility-checklist overstatements (e.g., claiming statistical tests without reporting them) are reporting issues, not circular reductions.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

No new physical or theoretical entities are postulated; the augmentation is a transformation of existing image statistics. The free parameters are limited to the manually chosen consistency weight λ, with no other data-fitted constants.

free parameters (1)
  • feature consistency weight λ = 0.17
    Hand-set in Eq. (12); no grid search or sensitivity analysis is reported, so the reported results may depend on this choice.
assumptions (5)
  • domain assumption RGB channel-wise mean and standard deviation carry domain characteristics, so perturbing them can generate samples from unseen domains.
    Stated in the Introduction and 'Data Augmentation in RGB Space'; it is the central premise of the augmentation. No empirical evidence is given within the paper to validate this specific mapping.
  • domain assumption Batch-level variances σ²_µc and σ²_σc estimated from a batch of size 4 provide a meaningful range for perturbation uncertainty.
    Equations (5)-(6) and the surrounding text; with B=4 the variance estimates are noisy, and the paper does not analyze sensitivity to batch size.
  • domain assumption Perturbing global per-channel statistics preserves scene structure and geometric content.
    Claimed under 'Structural Advantages'; needed so the same disparity ground truth remains valid for augmented images. The paper does not test this explicitly.
  • domain assumption Minimizing Lcons encourages shortcuts-invariant structure-aware features rather than merely reducing feature magnitude.
    Section 'Feature Consistency Constraint'; the L2 penalty could in principle be satisfied by feature shrinkage or collapse, a degenerate solution not discussed.
  • standard math Standard deep stereo training pipeline (cost volume, 3D CNNs) is taken as given.
    The method is inserted into existing networks (PSMNet, GwcNet, CFNet); the theoretical properties of these backbones are not re-derived.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Domain Generalized Stereo Matching with Uncertainty-guided Data Augmentation." pith.science (2026). https://pith.science/paper/CAWOA4YC

@misc{pith2026250801303,
  author       = {Pith},
  title        = {Pith review of: Domain Generalized Stereo Matching with Uncertainty-guided Data Augmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CAWOA4YC}},
  note         = {Machine review of arXiv:2508.01303}
}
read the original abstract

State-of-the-art stereo matching (SM) models trained on synthetic data often fail to generalize to real data domains due to domain differences, such as color, illumination, contrast, and texture. To address this challenge, we leverage data augmentation to expand the training domain, encouraging the model to acquire robust cross-domain feature representations instead of domain-dependent shortcuts. This paper proposes an uncertainty-guided data augmentation (UgDA) method, which argues that the image statistics in RGB space (mean and standard deviation) carry the domain characteristics. Thus, samples in unseen domains can be generated by properly perturbing these statistics. Furthermore, to simulate more potential domains, Gaussian distributions founded on batch-level statistics are poposed to model the unceratinty of perturbation direction and intensity. Additionally, we further enforce feature consistency between original and augmented data for the same scene, encouraging the model to learn structure aware, shortcuts-invariant feature representations. Our approach is simple, architecture-agnostic, and can be integrated into any SM networks. Extensive experiments on several challenging benchmarks have demonstrated that our method can significantly improve the generalization performance of existing SM networks.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 30 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Cai, C.; Poggi, M.; Mattoccia, S.; and Mordohai, P. 2020. Matching-space stereo networks for cross-domain generalization. In 2020 International conference on 3D vision (3DV), 364--373. IEEE

  4. [4]

    Chang, J.-R.; and Chen, Y.-S. 2018. Pyramid stereo matching network. In Proceedings of the IEEE conference on computer vision and pattern recognition, 5410--5418

  5. [5]

    Chang, T.; Yang, X.; Zhang, T.; and Wang, M. 2023. Domain generalized stereo matching via hierarchical visual transformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 9559--9568

  6. [6]

    Chuah, W.; Tennakoon, R.; Hoseinnezhad, R.; Bab-Hadiashar, A.; and Suter, D. 2022. Itsa: An information-theoretic approach to automatic shortcut avoidance and domain generalization in stereo matching networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 13022--13032

  7. [7]

    E.; Tran, Q.-H.; Zia, M

    Fathy, M. E.; Tran, Q.-H.; Zia, M. Z.; Vernaza, P.; and Chandraker, M. 2018. Hierarchical metric learning and matching for 2d and 3d geometric correspondences. In Proceedings of the european conference on computer vision (ECCV), 803--819

  8. [8]

    Geiger, A.; Lenz, P.; and Urtasun, R. 2012. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, 3354--3361. IEEE

Show all 34 references
  1. [9]

    Gu, X.; Fan, Z.; Zhu, S.; Dai, Z.; Tan, F.; and Tan, P. 2020. Cascade cost volume for high-resolution multi-view stereo and stereo matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2495--2504

  2. [10]

    Guo, X.; Yang, K.; Yang, W.; Wang, X.; and Li, H. 2019. Group-wise correlation stereo network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 3273--3282

  3. [11]

    S.; Abd Manap, N.; Hamzah, R

    Hamid, M. S.; Abd Manap, N.; Hamzah, R. A.; and Kadmin, A. F. 2022. Stereo matching algorithm based on deep learning: A survey. Journal of King Saud University-Computer and Information Sciences, 34(5): 1663--1673

  4. [12]

    Li, X.; Dai, Y.; Ge, Y.; Liu, J.; Shan, Y.; and Duan, L.-Y. 2022. Uncertainty modeling for out-of-distribution generalization. arXiv preprint arXiv:2202.03958

  5. [13]

    Li, X.; Fan, Y.; Rao, Z.; Lv, G.; and Liu, S. 2021. Synthetic-to-real domain adaptation joint spatial feature transform for stereo matching. IEEE Signal Processing Letters, 29: 60--64

  6. [14]

    Liu, B.; Yu, H.; and Qi, G. 2022. Graftnet: Towards domain generalized stereo matching with a broad-spectrum and task-oriented feature. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 13012--13021

  7. [15]

    Liu, R.; Yang, C.; Sun, W.; Wang, X.; and Li, H. 2020. Stereogan: Bridging synthetic-to-real domain gap by joint optimization of domain translation and stereo matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 12757--12766

  8. [16]

    Mayer, N.; Ilg, E.; Hausser, P.; Fischer, P.; Cremers, D.; Dosovitskiy, A.; and Brox, T. 2016. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In Proceedings of the IEEE conference on computer vision and pattern recogniti...

  9. [17]

    Menze, M.; and Geiger, A. 2015. Object scene flow for autonomous vehicles. In Proceedings of the IEEE conference on computer vision and pattern recognition, 3061--3070

  10. [18]

    K.; Li, C.; and Srikanthan, T

    Miao, Y.; Wu, M.; Lam, S. K.; Li, C.; and Srikanthan, T. 2024. Hierarchical object-aware dual-level contrastive learning for domain generalized stereo matching. Advances in Neural Information Processing Systems, 37: 132050--132076

  11. [19]

    Pang, J.; Sun, W.; Yang, C.; Ren, J.; Xiao, R.; Zeng, J.; and Lin, L. 2018. Zoom and learn: Generalizing deep stereo matching to novel domains. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2070--2079

  12. [20]

    Poggi, M.; Pallotti, D.; Tosi, F.; and Mattoccia, S. 2019. Guided stereo matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 979--988

  13. [21]

    Rao, Z.; Xiong, B.; He, M.; Dai, Y.; He, R.; Shen, Z.; and Li, X. 2023. Masked representation learning for domain generalized stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 5435--5444

  14. [22]

    Scharstein, D.; Hirschm \"u ller, H.; Kitajima, Y.; Krathwohl, G.; Ne s i \'c , N.; Wang, X.; and Westling, P. 2014. High-resolution stereo datasets with subpixel-accurate ground truth. In German conference on pattern recognition, 31--42. Springer

  15. [23]

    L.; Galliani, S.; Sattler, T.; Schindler, K.; Pollefeys, M.; and Geiger, A

    Schops, T.; Schonberger, J. L.; Galliani, S.; Sattler, T.; Schindler, K.; Pollefeys, M.; and Geiger, A. 2017. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, 3260--3269

  16. [24]

    Shen, Z.; Dai, Y.; and Rao, Z. 2021. Cfnet: Cascade and fused cost volume for robust stereo matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 13906--13915

  17. [25]

    Song, X.; Zhao, X.; Fang, L.; Hu, H.; and Yu, Y. 2020. Edgestereo: An effective multi-task learning network for stereo matching and edge detection. International Journal of Computer Vision, 128(4): 910--930

  18. [26]

    Wu, R.; Wang, M.; Li, Z.; Zhou, J.; Chen, F.; Wang, X.; and Sun, C. 2024. Few-shot stereo matching with high domain adaptability based on adaptive recursive network. International Journal of Computer Vision, 132(5): 1484--1501

  19. [27]

    Yang, G.; Manela, J.; Happold, M.; and Ramanan, D. 2019. Hierarchical deep stereo matching on high-resolution images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 5515--5524

  20. [28]

    Yang, G.; Zhao, H.; Shi, J.; Deng, Z.; and Jia, J. 2018. Segstereo: Exploiting semantic information for disparity estimation. In Proceedings of the European conference on computer vision (ECCV), 636--651

  21. [29]

    Yao, C.; Yu, L.; Liu, Z.; Zeng, J.; Wu, Y.; and Jia, Y. 2025. Diving into the Fusion of Monocular Priors for Generalized Stereo Matching. arXiv preprint arXiv:2505.14414

  22. [30]

    Z bontar, J.; and LeCun, Y. 2016. Stereo matching by training a convolutional neural network to compare image patches. Journal of Machine Learning Research, 17(65): 1--32

  23. [31]

    Zhang, F.; Prisacariu, V.; Yang, R.; and Torr, P. H. 2019. Ga-net: Guided aggregation net for end-to-end stereo matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 185--194

  24. [32]

    Zhang, F.; Qi, X.; Yang, R.; Prisacariu, V.; Wah, B.; and Torr, P. 2020. Domain-invariant stereo matching networks. In Computer Vision--ECCV 2020: 16th European Conference, Glasgow, UK, August 23--28, 2020, Proceedings, Part II 16, 420--439. Springer

  25. [33]

    Zhang, J.; Wang, X.; Bai, X.; Wang, C.; Huang, L.; Chen, Y.; Gu, L.; Zhou, J.; Harada, T.; and Hancock, E. R. 2022. Revisiting domain generalized stereo matching networks from a feature consistency perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and P...

  26. [34]

    Zheng, D.; Wu, X.-M.; Liu, Z.; Meng, J.; and Zheng, W.-s. 2025. Diffuvolume: Diffusion model for volume based stereo matching. International Journal of Computer Vision, 1--15

Pith tools

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