Pith. sign in

REVIEW 3 major objections 7 minor 2 cited by

Diving into the Fusion of Monocular Priors for Generalized Stereo Matching

T0 review · 3 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A stereo matching network trained only on synthetic data generalizes to real-world ill-posed regions by converting both monocular depth and disparity into binary local ordering maps and registering the monocular depth globally.

desk verdict A genuinely new fusion mechanism for monocular priors in stereo matching, well evaluated on zero-shot benchmarks, but the 'SOTA' claim is unanchored by the omission of the very concurrent methods the paper cites. read the letter →

arxiv 2505.14414 v2 pith:5DX3OPQA submitted 2025-05-20 cs.CV

Chengtang Yao , Lidong Yu , Zhidan Liu , Jiaxi Zeng , Yuwei Wu , Yunde Jia This is my paper · ORCID
classification cs.CV
keywords stereomatchingdomaingeneralizationmonoculardepthfusionbinarylocalorderingmapvisionfoundationmodelsill-posedregionszero-shotregistration
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 the poor generalization of stereo matching to real-world ill-posed regions—occlusions, textureless surfaces, reflections, and transparency—can be fixed by fusing an unbiased monocular prior from a large pretrained monocular model, provided the fusion is done the right way. The proposed way is to convert both the affine-invariant monocular depth and the absolute disparity into a binary local ordering map, so the two modalities first agree on who is closer than whom; only afterward does a global registration module learn per-pixel scale and shift to bring the monocular depth into the disparity's absolute range. Training only on the synthetic SceneFlow dataset, the method reports error reductions of more than half on two standard real-world stereo benchmarks and more than a 10-point improvement in transparent regions on the Booster benchmark, without extra stereo data or augmentation, while keeping runtime near that of the baseline. A sympathetic reading is that the ordering map is the essential step: it makes a strong prior usable without paying the usual alignment cost.

What carries the argument

The load-bearing mechanism is the binary local ordering map, defined for each pixel as the sigmoid of the depth difference between the center and its neighbors: $M_O(u,v) = \{\sigma(D(u',v')-D(u,v)) : (u',v')\in \mathcal{N}(u,v)\}$. Because it records only order, it is invariant to the affine gap between monocular depth and disparity, converting both into a common 'closer/farther' language. The second mechanism is the global registration module, which treats the optimized disparity as a noisy affine version of monocular depth and predicts pixel-wise scale $a$ and shift $b$; the aligned monocular depth is then blended with the disparity using a learned confidence map. A supporting detail is the gradual-release schedule $\tilde{\Delta d} = \Delta d (1 + G\cdot r\cdot t/T)$, which prevents noisy early-iteration disparity updates from corrupting the guidance.

What would settle it

On a real-world stereo pair with ground-truth disparity, fit the per-pixel affine map that best takes monocular depth to the final optimized disparity and check whether the residual after this affine fit is actually small enough to explain the reported error. If the residual is dominated by non-affine structure, the paper's registration model is not the mechanism responsible for the gains.

Watch

Extended reading notes

Core claim

The central claim is that a binary local ordering map $M_O$—computed from monocular depth and from the current disparity estimate via fixed LBP-like convolutions, storing only whether neighboring pixels are closer or farther—unifies relative and absolute depth representations and can act as a controllable guide for iterative disparity updates. The paper claims that re-weighting each disparity update by this guidance, with a gradual release over iterations to blunt early noise, avoids the local optima that naive monocular-feature fusion falls into. It further claims that the final fusion of monocular depth into disparity is a registration problem: a pixel-wise linear regression network $a,b = F(D_m, D_d^T)$ aligns the two, and a confidence map from cost volume, hidden state, and guidance blends them. The reported result is that these mechanisms, trained on SceneFlow alone, halve errors on two real-world benchmarks and improve transparent-region metrics by over 10 points compared with previous methods.

Load-bearing premise

The registration network and the ordering-map guidance are learned only on synthetic data and must transfer to real scenes with different disparity scales, materials, and scene statistics; if they overfit to the synthetic relationship between monocular depth and disparity, the reported zero-shot gains will vanish.

Editorial extensions

If this is right

  • Zero-shot deployment: a stereo network trained on synthetic data alone can be applied to real-world scenes with occlusions, reflections, and transparent surfaces without retraining or augmentation.
  • Efficiency: the generalization gain can be added on top of an iterative stereo backbone with a modest runtime increase (0.32s to 0.4s in the paper's measurements).
  • Ill-posed regions benefit most: the largest reported gains occur precisely where stereo matching is weakest, indicating the monocular prior supplies complementary shape information rather than simply sharpening already-good estimates.
  • No fixed disparity range: because the fusion is an affine registration rather than a volume injection, the method does not assume a predefined disparity range, which helps high-resolution and large-baseline inputs.
  • Modularity: the monocular encoder and the fusion modules are separable, so better stereo backbones or better monocular priors could be swapped in without redesigning the fusion.

Reading between the lines

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

  • Not tested in the paper: whether the same binary-ordering fusion helps when the monocular prior is weaker or noisier, such as a smaller pretrained model; the paper's mechanism suggests robustness, but that is an inference from the ablation, not a demonstrated result.
  • A concrete follow-up experiment would train the same modules with the monocular encoder frozen at a much lower input resolution to test whether the ordering map's scale invariance protects against resolution loss.
  • The failure cases the paper lists—glass doors and dark tunnels—suggest a next step is representing two depths per pixel; the binary ordering map would need to become a multi-label ordering, which the paper only mentions as future work.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 7 minor

Summary. The paper presents a stereo matching network trained on SceneFlow and designed to generalize zero-shot to real-world datasets by fusing monocular depth priors from a frozen DepthAnything V2 model. The method has three main components: a monocular encoder that extracts context features and depth, an iterative local fusion module that uses a binary local ordering map to re-weight the disparity update, and a global fusion module that registers the monocular depth to the optimized disparity via per-pixel affine parameters and blends them with a learned confidence map. The paper evaluates on KITTI 2012/2015, Middlebury, ETH3D, Booster, and DrivingStereo, reporting large improvements over stereo-only baselines and over its own ablations.

Significance. The paper's central contribution is a principled way to inject relative monocular depth into an iterative stereo optimizer without fine-tuning the VFM. The binary local ordering map is an elegant device for unifying relative monocular depth and absolute disparity, and the per-pixel affine registration with confidence-based fusion is simple and efficient. The empirical protocol is mostly sound: standard training-on-SceneFlow evaluation, all-region metrics, re-evaluation of several baselines with official weights, and ablations reported as mean±std over training steps. The fixed-weight LBP-like operations are a thoughtful choice to avoid overfitting on small training data. However, the absence of any comparison to the three cited monocular-fusion methods (FoundationStereo, Monster, Stereo Anywhere) leaves the headline generalization claim unsubstantiated relative to the closest prior art.

major comments (3)
  1. [Sec. 2.2, Tables 1-3] The Related Work (§2.2) names FoundationStereo [50], Monster [8], and Stereo Anywhere [3] as recent stereo methods that fuse unbiased VFM monocular priors, yet none of these systems appears in any of the experimental tables (Tables 1-3). Because the central claim is that the proposed fusion modules generalize substantially better than prior stereo-only and monocular-fusion approaches, the absence of these direct competitors leaves the headline 'SOTA' comparison unanchored. The authors should add these baselines under the same all-region, zero-shot protocol, or explicitly re-scope the claim to stereo-only methods and justify why the cited concurrent work is not directly comparable.
  2. [Sec. 4.3, Table 3] The text in §4.3 states 'our method achieves state-of-the-art results across all datasets,' but Table 3 shows NerfStereo (with additional data) outperforms Ours on DrivingStereo in Sunny (0.90 vs 0.93), Cloudy (0.91 vs 0.92), and Foggy (1.01 vs 0.93). If the SOTA claim is intended only among methods without additional data, that restriction is not stated and the wording should be adjusted; if the claim includes all methods, it is contradicted by the paper's own table.
  3. [Conclusion, Table 1] The conclusion says 'an error reduction of more than half on Middlebury and ETH3D.' In Table 1, the reductions from the best no-extra-data baseline (IGEV) are: Middlebury EPE 2.27→0.85 (62% reduction) but Bad 2.0 9.49→5.67 (40% reduction); ETH3D EPE 0.33→0.25 (24% reduction) but Bad 2.0 4.0→1.88 (53% reduction). The paper should specify the metric and dataset pairing for this claim, since no single metric gives 'more than half' on both datasets.
minor comments (7)
  1. [Fig. 2, Sec. 4.1] The notation in Figure 2 uses circled symbols for warp and lookup operations, but these are not standard and may be difficult to parse; please use conventional labels and fix the missing space in 'configured to5, 3'.
  2. [Eq. (1)] The neighborhood N(u,v) in Eq. (1) is not defined; please state its size and whether the sigmoid is applied per neighbor difference.
  3. [Table 5] The caption lists 'DF' while the text refers to 'IDF' for iterative direct fusion; please make the abbreviations consistent.
  4. [Table 6] Row L(9) contains a trailing comma in the LBP kernel column; please correct the typo for consistency with the other rows.
  5. [Conclusion] The first sentence of the final paragraph contains a duplicated 'and' ('to extract ... and and leveraged'), and 'VTF' should be 'VFM' to match the rest of the paper.
  6. [Sec. 4.4] The runtime comparison states the baseline takes 0.32s and the model takes 0.4s, which is a 25% increase; the abstract's 'barely reducing the efficiency' should be qualified or the numbers should be rechecked.
  7. [Sec. 4.2] The paper emphasizes mean±std due to training fluctuations, but Tables 1-3 report single values; please clarify whether these are from a single checkpoint or averaged over steps.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation found: the method is trained on SceneFlow and evaluated zero-shot on external benchmarks; the only self-citation is a supplementary, non-load-bearing scalability experiment.

full rationale

The paper's derivation chain is self-contained rather than circular. The two novel components—the binary local ordering map (Eqs. 1–3, supervised through the disparity-update loss in Eq. 6) and the global registration module (Eqs. 4–5, supervised by its own L1 terms in Eq. 6)—are learned only on synthetic SceneFlow and then evaluated zero-shot on KITTI, Middlebury, ETH3D, Booster, and DrivingStereo. No test-set statistic is used to fit a parameter: the registration scale a and shift b are predictions of a trained network F, not a least-squares fit to the test disparity, and the confidence c is likewise learned. The only self-citation is reference [24] (TranScene) in the supplementary Section 9, used as an optional additional-training-data experiment; it plays no role in the main generalization claim, so it is not a load-bearing circular step. The absence of FoundationStereo, Monster, and Stereo Anywhere from the comparison tables is a legitimate experimental-coverage concern, and the '10-point improvement' and 'error reduction of more than half' claims are therefore unanchored against those closest monocular-fusion baselines. However, a missing comparison is not a circular derivation: it does not make the output equivalent to the input by construction, and no quoted equation reduces to the paper's own assumptions. Accordingly, no circular step meeting the evidence bar can be exhibited.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the transferability of learned modules (ordering-map guidance and per-pixel affine registration) from synthetic SceneFlow to real domains, and on the assumption that the frozen monocular foundation model supplies useful priors. The manually chosen free parameters are hyperparameters with reported low sensitivity. No new physical or theoretical entities are introduced.

free parameters (4)
  • Guidance amplitude r = 1
    Hand-chosen in Eq. 2, it scales the influence of the monocular guidance on the disparity update. Ablation in Table 6 shows mild sensitivity when varied between 1 and 3.
  • Loss balancing scalar gamma = 0.9
    Hand-chosen in Eq. 6, it weights the L1 losses over the iterations of disparity updates.
  • LBP-like window sizes = [5, 3]
    Chosen by ablation in Table 6 among several configurations for the fixed-weight local ordering feature extractor.
  • Training stage learning rates = 2e-4 / 5e-4
    Used in the three-stage training protocol on SceneFlow described in Sec. 4.1; selected manually and not reported as sensitive.
assumptions (5)
  • domain assumption The frozen DepthAnything V2 model provides an unbiased, domain-generalizable monocular prior.
    Sec. 3.1 states that the monocular priors learned by the stereo model are biased and that DepthAnything V2 is used to mitigate the domain gap. The entire method depends on this prior being reliable on unseen real scenes.
  • ad hoc to paper Neighboring relative depth ordering is preserved across domains and is more robust to noise than raw depth values.
    Eq. 1 and Sec. 3.2 motivate the binary local ordering map with this claim, which is not independently proven beyond the empirical results.
  • ad hoc to paper The optimized disparity can be modeled as a per-pixel affine registration of monocular depth with small noise.
    Eq. 4 and Sec. 3.3 formulate disparity as a noisy linear regression of monocular depth with pixel-wise scale and shift; this is the core assumption of the global fusion module.
  • ad hoc to paper A confidence map estimated from the cost volume, hidden state, and guidance can correctly blend stereo disparity and registered monocular depth.
    Sec. 3.3 predicts confidence with a network trained on SceneFlow and assumes it transfers to real-world scenes, supported only by the ablation in Table 7.
  • standard math Standard backpropagation and reparameterized Beta sampling are valid for training the stochastic guidance variable.
    Sec. 3.2 uses a Beta distribution with reparameterization; this is standard practice and not in question.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Diving into the Fusion of Monocular Priors for Generalized Stereo Matching." pith.science (2026). https://pith.science/paper/5DX3OPQA

@misc{pith2026250514414,
  author       = {Pith},
  title        = {Pith review of: Diving into the Fusion of Monocular Priors for Generalized Stereo Matching},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5DX3OPQA}},
  note         = {Machine review of arXiv:2505.14414}
}
read the original abstract

The matching formulation makes it naturally hard for the stereo matching to handle ill-posed regions like occlusions and non-Lambertian surfaces. Fusing monocular priors has been proven helpful for ill-posed matching, but the biased monocular prior learned from small stereo datasets constrains the generalization. Recently, stereo matching has progressed by leveraging the unbiased monocular prior from the vision foundation model (VFM) to improve the generalization in ill-posed regions. We dive into the fusion process and observe three main problems limiting the fusion of the VFM monocular prior. The first problem is the misalignment between affine-invariant relative monocular depth and absolute depth of disparity. Besides, when we use the monocular feature in an iterative update structure, the over-confidence in the disparity update leads to local optima results. A direct fusion of a monocular depth map could alleviate the local optima problem, but noisy disparity results computed at the first several iterations will misguide the fusion. In this paper, we propose a binary local ordering map to guide the fusion, which converts the depth map into a binary relative format, unifying the relative and absolute depth representation. The computed local ordering map is also used to re-weight the initial disparity update, resolving the local optima and noisy problem. In addition, we formulate the final direct fusion of monocular depth to the disparity as a registration problem, where a pixel-wise linear regression module can globally and adaptively align them. Our method fully exploits the monocular prior to support stereo matching results effectively and efficiently. We significantly improve the performance from the experiments when generalizing from SceneFlow to Middlebury and Booster datasets while barely reducing the efficiency.

Figures

Figures reproduced from arXiv: 2505.14414 by the authors.

Figure 1
Figure 1. The visualization of different ill-posed regions in the Booster dataset. Our method achieves an overwhelming advantage in all [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The pipeline of our method. ⃝w represents the warping operation when constructing the cost volume. ⃝L is look up operation used to sample cost volume. ⃝c represents concatenation. ⃝+ represents add operation, while ⃝× represents multiplication. bors: M_O(u,v) = \{ \sigma (D(u',v') - D(u, v)) \}, (1) where (u ′ , v′ ) ∈ N(u,v) , \sigma is the sigmoid function and \protect \mathcal {N} is the neighborhood. The binary … view at source ↗
Figure 3
Figure 3. The visualization of local ordering map. The monocular [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (20 more)
Figure 4
Figure 4. Figure 4: The visualization of registration parameters, scale [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: The visualization of results on the Flicker1024 dataset. [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: The visualization of results. We use two kinds of col [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: The visualization of binocular disparity and monocular depth. The regions highlighted with gray boxes demonstrate that stereo [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]
Figure 8
Figure 8. Figure 8: The visualization of binocular disparity and monocular depth. The regions highlighted with gray boxes demonstrate that stereo [PITH_FULL_IMAGE:figures/full_fig_p018_8.png]
Figure 9
Figure 9. Figure 9: The visualization for generalized stereo matching. [PITH_FULL_IMAGE:figures/full_fig_p019_9.png]
Figure 10
Figure 10. Figure 10: The visualization for generalized stereo matching. [PITH_FULL_IMAGE:figures/full_fig_p020_10.png]
Figure 11
Figure 11. Figure 11: The visualization for generalized stereo matching. [PITH_FULL_IMAGE:figures/full_fig_p021_11.png]
Figure 12
Figure 12. Figure 12: The visualization for generalized stereo matching. [PITH_FULL_IMAGE:figures/full_fig_p022_12.png]
Figure 13
Figure 13. Figure 13: The visualization for generalized stereo matching. [PITH_FULL_IMAGE:figures/full_fig_p023_13.png]
Figure 14
Figure 14. Figure 14: The visualization of intermediate results. [PITH_FULL_IMAGE:figures/full_fig_p024_14.png]
Figure 15
Figure 15. Figure 15: The visualization of intermediate results. [PITH_FULL_IMAGE:figures/full_fig_p025_15.png]
Figure 16
Figure 16. Figure 16: The visualization of intermediate results. [PITH_FULL_IMAGE:figures/full_fig_p026_16.png]
Figure 17
Figure 17. Figure 17: The visualization of intermediate results. [PITH_FULL_IMAGE:figures/full_fig_p027_17.png]
Figure 18
Figure 18. Figure 18: The visualization of intermediate results. [PITH_FULL_IMAGE:figures/full_fig_p028_18.png]
Figure 19
Figure 19. Figure 19: The visualization of intermediate results. [PITH_FULL_IMAGE:figures/full_fig_p029_19.png]
Figure 20
Figure 20. Figure 20: The visualization of intermediate results. [PITH_FULL_IMAGE:figures/full_fig_p030_20.png]
Figure 21
Figure 21. Figure 21: The visualization of intermediate results. [PITH_FULL_IMAGE:figures/full_fig_p030_21.png]
Figure 22
Figure 22. Figure 22: The visualization for failure case analysis. [PITH_FULL_IMAGE:figures/full_fig_p031_22.png]
Figure 23
Figure 23. Figure 23: The visualization for failure case analysis. [PITH_FULL_IMAGE:figures/full_fig_p032_23.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. STEREOFLOW: Progressive Stereo Matching with StereoDiT and Transition Flow Matching

    cs.CV 2026-07 conditional novelty 5.0 of 10

    A hybrid stereo-matching model uses a cascade matching network to propose disparities and a diffusion transformer to refine ambiguous regions; it claims state-of-the-art benchmark results.

  2. Domain Generalized Stereo Matching with Uncertainty-guided Data Augmentation

    cs.CV 2025-08 conditional novelty 4.0 of 10

    UgDA-Stereo improves cross-domain stereo matching by randomly perturbing per-channel image mean and variance, guided by batch-level statistics, plus a feature consistency loss.

Reference graph

Works this paper leans on

68 extracted references · 63 canonical work pages · cited by 2 Pith papers

  1. [50]

    Foundationstereo: Zero- shot stereo matching

    Bowen Wen, Matthew Trepte, Joseph Aribido, Jan Kautz, Orazio Gallo, and Stan Birchfield. Foundationstereo: Zero- shot stereo matching. arXiv preprint arXiv:2501.09898 ,

    arXiv

  2. [8]

    Monster: Marry monodepth to stereo unleashes power

    Junda Cheng, Longliang Liu, Gangwei Xu, Xianqi Wang, Zhaoxing Zhang, Yong Deng, Jinliang Zang, Yurui Chen, Zhipeng Cai, and Xin Yang. Monster: Marry monodepth to stereo unleashes power. arXiv preprint arXiv:2501.08643 ,

    arXiv

  3. [3]

    Stereo anywhere: Robust zero-shot deep stereo matching even where either stereo or mono fail

    Luca Bartolomei, Fabio Tosi, Matteo Poggi, and Stefano Mattoccia. Stereo anywhere: Robust zero-shot deep stereo matching even where either stereo or mono fail. arXiv preprint arXiv:2412.04472, 2024. 1, 3

    arXiv 2024

  4. [1]

    Reversing the cycle: Self-supervised deep stereo through enhanced monocular distillation

    Filippo Aleotti, Fabio Tosi, Li Zhang, Matteo Poggi, and Ste- fano Mattoccia. Reversing the cycle: Self-supervised deep stereo through enhanced monocular distillation. InEuropean Conference on Computer Vision, pages 614–632, 2020. 3

    work page 2020

  5. [2]

    Multi- view depth estimation by fusing single-view depth prob- ability with multi-view geometry

    Gwangbin Bae, Ignas Budvytis, and Roberto Cipolla. Multi- view depth estimation by fusing single-view depth prob- ability with multi-view geometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2842–2851, 2022. 3

    work page 2022

  6. [4]

    Matching-space stereo networks for cross-domain generalization

    Changjiang Cai, Matteo Poggi, Stefano Mattoccia, and Philippos Mordohai. Matching-space stereo networks for cross-domain generalization. In Proceedings of the Interna- tional Conference on 3D Vision, pages 364–373. IEEE, 2020. 3

    work page 2020

  7. [5]

    Domain generalized stereo matching via hierarchical visual transformation

    Tianyu Chang, Xun Yang, Tianzhu Zhang, and Meng Wang. Domain generalized stereo matching via hierarchical visual transformation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages 9559– 9568, 2023. 3, 6

    work page 2023

  8. [6]

    Revealing the reciprocal relations between self- supervised stereo and monocular depth estimation

    Zhi Chen, Xiaoqing Ye, Wei Yang, Zhenbo Xu, Xiao Tan, Zhikang Zou, Errui Ding, Xinming Zhang, and Liusheng Huang. Revealing the reciprocal relations between self- supervised stereo and monocular depth estimation. In Pro- ceedings of the IEEE International Conference on Computer Vision, pages 15529–15538, 2021. 3

    work page 2021

Show all 68 references
  1. [7]

    Mocha-stereo: Motif channel attention network for stereo matching

    Ziyang Chen, Wei Long, He Yao, Yongjun Zhang, Bing- shu Wang, Yongbin Qin, and Jia Wu. Mocha-stereo: Motif channel attention network for stereo matching. In Proceed- ings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 27768–27777, 2024. 1, 3, 6, 2, 5

    2024

  2. [9]

    Itsa: An information-theoretic approach to automatic shortcut avoid- ance and domain generalization in stereo matching networks

    WeiQin Chuah, Ruwan Tennakoon, Reza Hoseinnezhad, Alireza Bab-Hadiashar, and David Suter. Itsa: An information-theoretic approach to automatic shortcut avoid- ance and domain generalization in stereo matching networks. In Proceedings of the IEEE Conference on Computer Vision a...

    2022

  3. [10]

    Perceiving layout and knowing distances: The integration, relative potency, and contextual use of different information about depth

    James E Cutting and Peter M Vishton. Perceiving layout and knowing distances: The integration, relative potency, and contextual use of different information about depth. In Per- ception of space and motion , pages 69–117. Elsevier, 1995. 3

    1995

  4. [11]

    Are we ready for autonomous driving? the kitti vision benchmark suite

    Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Conference on Computer Vision and Pattern Recog- nition (CVPR), 2012. 5

    2012

  5. [12]

    Neural markov random field for stereo matching

    Tongfan Guan, Chen Wang, and Yun-Hui Liu. Neural markov random field for stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459–5469, 2024. 3, 6

    2024

  6. [13]

    On calibration of modern neural networks

    Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In International Conference on Machine Learning, pages 1321–1330. PMLR,

  7. [14]

    Context-enhanced stereo transformer

    Weiyu Guo, Zhaoshuo Li, Yongkui Yang, Zheng Wang, Rus- sell H Taylor, Mathias Unberath, Alan Yuille, and Yingwei Li. Context-enhanced stereo transformer. In Proceedings of the European Conference on Computer Vision , pages 263–

  8. [15]

    Uncertainty guided adaptive warping for robust and efficient stereo matching

    Junpeng Jing, Jiankun Li, Pengfei Xiong, Jiangyu Liu, Shuaicheng Liu, Yichen Guo, Xin Deng, Mai Xu, Lai Jiang, and Leonid Sigal. Uncertainty guided adaptive warping for robust and efficient stereo matching. In Proceedings of the IEEE International Conference on Computer Vision...

    2023

  9. [16]

    Repurpos- ing diffusion-based image generators for monocular depth estimation

    Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Met- zger, Rodrigo Caye Daudt, and Konrad Schindler. Repurpos- ing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 9492...

    2024

  10. [17]

    Pointfix: Learning to fix domain bias for robust online stereo adaptation

    Kwonyoung Kim, Jungin Park, Jiyoung Lee, Dongbo Min, and Kwanghoon Sohn. Pointfix: Learning to fix domain bias for robust online stereo adaptation. In European Conference on Computer Vision, pages 568–585. Springer, 2022. 3

    2022

  11. [18]

    Ground- ing image matching in 3d with mast3r, 2024

    Vincent Leroy, Yohann Cabon, and Jerome Revaud. Ground- ing image matching in 3d with mast3r, 2024. 3

    2024

  12. [19]

    Los: Local structure-guided stereo matching

    Kunhong Li, Longguang Wang, Ye Zhang, Kaiwen Xue, Shunbo Zhou, and Yulan Guo. Los: Local structure-guided stereo matching. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition , pages 19746–19756, 2024. 1

    2024

  13. [20]

    Learning to fuse monocular and multi-view cues for multi- frame depth estimation in dynamic scenes

    Rui Li, Dong Gong, Wei Yin, Hao Chen, Yu Zhu, Kaix- uan Wang, Xiaozhi Chen, Jinqiu Sun, and Yanning Zhang. Learning to fuse monocular and multi-view cues for multi- frame depth estimation in dynamic scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern R...

    2023

  14. [21]

    Revisiting stereo depth estimation from a sequence- to-sequence perspective with transformers

    Zhaoshuo Li, Xingtong Liu, Nathan Drenkow, Andy Ding, Francis X Creighton, Russell H Taylor, and Mathias Un- berath. Revisiting stereo depth estimation from a sequence- to-sequence perspective with transformers. In Proceedings of the IEEE International Conference on Computer V...

    2021

  15. [22]

    Raft-stereo: Multilevel recurrent field transforms for stereo matching

    Lahav Lipson, Zachary Teed, and Jia Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. In 2021 International Conference on 3D Vision , pages 218–

    2021

  16. [23]

    Graftnet: Towards domain generalized stereo matching with a broad-spectrum and task-oriented feature

    Biyang Liu, Huimin Yu, and Guodong Qi. Graftnet: Towards domain generalized stereo matching with a broad-spectrum and task-oriented feature. In Proceedings of the IEEE Con- ference on Computer Vision and Pattern Recognition, pages 13012–13021, 2022. 1, 3, 6

    2022

  17. [24]

    Multi-label stereo matching for transparent scene depth estimation

    Zhidan Liu, Chengtang Yao, Jiaxi Zeng, Yuwei Wu, and Yunde Jia. Multi-label stereo matching for transparent scene depth estimation. ArXiv, 2025. 2

    2025

  18. [25]

    Elfnet: Evidential local-global fusion for stereo matching

    Jieming Lou, Weide Liu, Zhuo Chen, Fayao Liu, and Jun Cheng. Elfnet: Evidential local-global fusion for stereo matching. In Proceedings of the IEEE International Con- ference on Computer Vision, pages 17784–17793, 2023. 1, 3, 6

    2023

  19. [26]

    Fu- sion of stereo and still monocular depth estimates in a self- supervised learning context

    Diogo Martins, Kevin Van Hecke, and Guido De Croon. Fu- sion of stereo and still monocular depth estimates in a self- supervised learning context. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 849–

    2018

  20. [27]

    A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation

    Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. InProceedings of the IEEE Conference on Computer Vision and P...

    2016

  21. [28]

    Joint 3d estimation of vehicles and scene flow

    Moritz Menze, Christian Heipke, and Andreas Geiger. Joint 3d estimation of vehicles and scene flow. InISPRS Workshop on Image Sequence Analysis (ISA), 2015. 5

    2015

  22. [29]

    Ob- ject scene flow

    Moritz Menze, Christian Heipke, and Andreas Geiger. Ob- ject scene flow. ISPRS Journal of Photogrammetry and Re- mote Sensing (JPRS), 2018. 5

    2018

  23. [30]

    Per- formance evaluation of texture measures with classification based on kullback discrimination of distributions

    Timo Ojala, Matti Pietikainen, and David Harwood. Per- formance evaluation of texture measures with classification based on kullback discrimination of distributions. In Pro- ceedings of 12th international conference on pattern recog- nition, pages 582–585. IEEE, 1994. 4

    1994

  24. [31]

    Multires- olution gray-scale and rotation invariant texture classification with local binary patterns

    Timo Ojala, Matti Pietikainen, and Topi Maenpaa. Multires- olution gray-scale and rotation invariant texture classification with local binary patterns. IEEE Transactions on pattern analysis and machine intelligence, 24(7):971–987, 2002. 4

    2002

  25. [32]

    Can you trust your model’s uncertainty? evaluating predictive uncertainty under dataset shift

    Yaniv Ovadia, Emily Fertig, Jie Ren, Zachary Nado, David Sculley, Sebastian Nowozin, Joshua Dillon, Balaji Lakshmi- narayanan, and Jasper Snoek. Can you trust your model’s uncertainty? evaluating predictive uncertainty under dataset shift. Advances in neural information proces...

    2019

  26. [33]

    Open chal- lenges in deep stereo: the booster dataset

    Pierluigi Zama Ramirez, Fabio Tosi, Matteo Poggi, Samuele Salti, Stefano Mattoccia, and Luigi Di Stefano. Open chal- lenges in deep stereo: the booster dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 21168–21178, 2022. 5

    2022

  27. [34]

    Vi- sion transformers for dense prediction

    Ren ´e Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vi- sion transformers for dense prediction. In Proceedings of the IEEE international conference on computer vision, pages 12179–12188, 2021. 3

    2021

  28. [35]

    Masked representation learning for domain generalized stereo matching

    Zhibo Rao, Bangshu Xiong, Mingyi He, Yuchao Dai, Ren- jie He, Zhelun Shen, and Xing Li. Masked representation learning for domain generalized stereo matching. In Pro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5435–5444, 2023. 1, 3, 6

    2023

  29. [36]

    The perception of egocentric distances in virtual environments-a review

    Rebekka S Renner, Boris M Velichkovsky, and Jens R Helmert. The perception of egocentric distances in virtual environments-a review. ACM Computing Surveys (CSUR) , 46(2):1–40, 2013. 3

    2013

  30. [37]

    Depth estimation using monocular and stereo cues

    Ashutosh Saxena, Jamie Schulte, Andrew Y Ng, et al. Depth estimation using monocular and stereo cues. InInternational Joint Conference on Artificial Intelligence (IJCAI) , pages 2197–2203, 2007. 3

    2007

  31. [38]

    High-resolution stereo datasets with subpixel-accurate ground truth

    Daniel Scharstein, Heiko Hirschm ¨uller, York Kitajima, Greg Krathwohl, Nera Ne ˇsi´c, Xi Wang, and Porter West- ling. High-resolution stereo datasets with subpixel-accurate ground truth. In German Conference on Pattern Recogni- tion), pages 31–42, 2014. 5

    2014

  32. [39]

    A multi-view stereo benchmark with high- resolution images and multi-camera videos

    Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and An- dreas Geiger. A multi-view stereo benchmark with high- resolution images and multi-camera videos. In Proceedings of the IEEE Conference on Computer Vision and Pa...

    2017

  33. [40]

    Pcw-net: Pyramid combination and warping cost volume for stereo matching

    Zhelun Shen, Yuchao Dai, Xibin Song, Zhibo Rao, Dingfu Zhou, and Liangjun Zhang. Pcw-net: Pyramid combination and warping cost volume for stereo matching. In Proceed- ings of the European Conference on Computer Vision, pages 280–297. Springer, 2022. 6

    2022

  34. [41]

    Adastereo: An efficient domain-adaptive stereo matching approach

    Xiao Song, Guorun Yang, Xinge Zhu, Hui Zhou, Yuexin Ma, Zhe Wang, and Jianping Shi. Adastereo: An efficient domain-adaptive stereo matching approach. International Journal of Computer Vision (IJCV), pages 1–20, 2022. 3

    2022

  35. [42]

    Learn- ing to adapt for stereo

    Alessio Tonioni, Oscar Rahnama, Thomas Joy, Luigi Di Ste- fano, Thalaiyasingam Ajanthan, and Philip HS Torr. Learn- ing to adapt for stereo. In Proceedings of the IEEE Con- ference on Computer Vision and Pattern Recognition, pages 9661–9670, 2019. 3

    2019

  36. [43]

    Nerf-supervised deep stereo

    Fabio Tosi, Alessio Tonioni, Daniele De Gregorio, and Mat- teo Poggi. Nerf-supervised deep stereo. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 855–866, 2023. 3, 6

    2023

  37. [44]

    Gated stereo: Joint depth estimation from gated and wide-baseline active stereo cues

    Stefanie Walz, Mario Bijelic, Andrea Ramazzina, Aman- preet Walia, Fahim Mannan, and Felix Heide. Gated stereo: Joint depth estimation from gated and wide-baseline active stereo cues. In Proceedings of the IEEE Conference on Com- puter Vision and Pattern Recognition , pages 13...

  38. [45]

    Dust3r: Geometric 3d vi- sion made easy

    Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vi- sion made easy. In CVPR, 2024. 3

    2024

  39. [46]

    Selective-stereo: Adaptive frequency information selection for stereo matching

    Xianqi Wang, Gangwei Xu, Hao Jia, and Xin Yang. Selective-stereo: Adaptive frequency information selection for stereo matching. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 19701– 19710, 2024. 1, 3, 6, 2, 5

    2024

  40. [47]

    Flickr1024: A large-scale dataset for stereo image super-resolution

    Yingqian Wang, Longguang Wang, Jungang Yang, Wei An, and Yulan Guo. Flickr1024: A large-scale dataset for stereo image super-resolution. In International Conference on Computer Vision Workshops, pages 3852–3857, 2019. 1

    2019

  41. [48]

    The human brain in depth: how we see in 3d

    Andrew E Welchman. The human brain in depth: how we see in 3d. Annual review of vision science , 2(1):345–376,

  42. [49]

    3d shape perception from combined depth cues in human visual cortex

    Andrew E Welchman, Arne Deubelius, Verena Conrad, Heinrich H B ¨ulthoff, and Zoe Kourtzi. 3d shape perception from combined depth cues in human visual cortex. Nature neuroscience, 8(6):820–827, 2005

    2005

  43. [51]

    Iterative geometry encoding volume for stereo matching

    Gangwei Xu, Xianqi Wang, Xiaohuan Ding, and Xin Yang. Iterative geometry encoding volume for stereo matching. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 21919–21928, 2023. 1, 3, 6, 2, 5

    2023

  44. [52]

    Depth anything: Un- leashing the power of large-scale unlabeled data

    Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Ji- ashi Feng, and Hengshuang Zhao. Depth anything: Un- leashing the power of large-scale unlabeled data. In Pro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 10371–10381, 2024. 1, 2, 5

    2024

  45. [53]

    Depth any- thing v2

    Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiao- gang Xu, Jiashi Feng, and Hengshuang Zhao. Depth any- thing v2. arXiv preprint arXiv:2406.09414 , 2024. 1, 3, 5, 2

    2024 arXiv

  46. [54]

    Metric3d: Towards zero-shot metric 3d prediction from a single image

    Wei Yin, Chi Zhang, Hao Chen, Zhipeng Cai, Gang Yu, Kaixuan Wang, Xiaozhi Chen, and Chunhua Shen. Metric3d: Towards zero-shot metric 3d prediction from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9043–9053, 2023. 2

    2023

  47. [55]

    Multi-view stereo by fusing monocular and a combination of depth representation meth- ods

    Fanqi Yu and Xinyang Sun. Multi-view stereo by fusing monocular and a combination of depth representation meth- ods. In International Conference on Neural Information Pro- cessing (NeurIPS), pages 298–309. Springer, 2023. 3

    2023

  48. [56]

    Stereo depth es- timation with echoes

    Chenghao Zhang, Kun Tian, Bolin Ni, Gaofeng Meng, Bin Fan, Zhaoxiang Zhang, and Chunhong Pan. Stereo depth es- timation with echoes. In European Conference on Computer Vision, pages 496–513. Springer, 2022. 3

    2022

  49. [57]

    Domain-invariant stereo matching networks

    Feihu Zhang, Xiaojuan Qi, Ruigang Yang, Victor Prisacariu, Benjamin Wah, and Philip Torr. Domain-invariant stereo matching networks. In Proceedings of the European Con- ference on Computer Vision, pages 420–439. Springer, 2020. 3

    2020

  50. [58]

    Revisiting domain generalized stereo matching networks from a feature consistency perspective

    Jiawei Zhang, Xiang Wang, Xiao Bai, Chen Wang, Lei Huang, Yimin Chen, Lin Gu, Jun Zhou, Tatsuya Harada, and Edwin R Hancock. Revisiting domain generalized stereo matching networks from a feature consistency perspective. In Proceedings of the IEEE Conference on Computer Vision ...

    2022

  51. [59]

    Two-in-one depth: Bridging the gap between monocular and binocular self- supervised depth estimation

    Zhengming Zhou and Qiulei Dong. Two-in-one depth: Bridging the gap between monocular and binocular self- supervised depth estimation. In Proceedings of the IEEE International Conference on Computer Vision, pages 9411– 9421, 2023. 3 Diving into the Fusion of Monocular Priors fo...

    2023

  52. [61]

    Visualization On Flicker1024 We present visualization results demonstrating the general- ization capability of our model from the synthetic Scene- Flow dataset to the real-world Flickr1024 dataset [47]. As shown in Figure 5, our model performs robustly across di- verse scenari...

  53. [62]

    As shown in Figure 6, DepthAnything v2 provides depth maps with better continuity than Marigold, especially in fine-grained regions

    Intuition behind Monocular Depth Model We choose DepthAnything v2 [53] over Marigold [16] be- cause of the superior continuity of its depth maps. As shown in Figure 6, DepthAnything v2 provides depth maps with better continuity than Marigold, especially in fine-grained regions...

  54. [63]

    From class 0 to 3, the materials be- come increasingly transparent and/or specular

    More Results on Booster We provide additional results on the Booster dataset across various material types. From class 0 to 3, the materials be- come increasingly transparent and/or specular. As shown in Tables 8 and 9, our method outperforms state-of-the-art ap- proaches on t...

  55. [64]

    In our experiments, we use labels with the largest disparity in transparent regions

    Additional Training Data We evaluate the scalability of our model by incorporating additional training data from the TranScene dataset [24], a synthetic dataset specifically designed for multi-label trans- parent scenes. In our experiments, we use labels with the largest dispa...

  56. [65]

    To ensure a fair comparison of backbones during inference, we exclude the feature encoder module when evaluating each model’s memory consumption

    More Analysis about Memory We also compare our model to state-of-the-art methods in terms of memory consumption across different resolutions. To ensure a fair comparison of backbones during inference, we exclude the feature encoder module when evaluating each model’s memory co...

  57. [66]

    More Visualization We provide additional visualizations of generalized stereo matching in Figures 9, 10, 11, 12, and 13. The visu- alizations span a variety of environments, ranging from open outdoor scenes (e.g., driving scenarios), to semi-open outdoor scenes (e.g., playgrou...

  58. [67]

    Ablation Study 12.1. More Analysis of Backbone In addition to replacing the context network with the pre- trained DepthAnything v2 [53], we also experimented with replacing the feature extractor for cost volume construction using DepthAnything v2 [53] and MASt3R [18, 45]. As s...

  59. [68]

    In the first failure case, our method is confused by the glass door and glass window, where both the transparent surfaces and the behind scene are significant

    Future Work Discussion We present failure cases in Figures 22 and 23. In the first failure case, our method is confused by the glass door and glass window, where both the transparent surfaces and the behind scene are significant. Unlike simple transparent ob- jects (e.g., a gl...

  60. [227]

    5, 6, 2, 3

    IEEE, 2021. 5, 6, 2, 3

    2021

Pith tools

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