Pith. sign in

REVIEW 3 major objections 5 minor 126 references

PanMatch: Unleashing the Potential of Large Vision Models for Unified Matching Models

T0 review · 3 major / 5 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 One model with identical weights can perform stereo matching, optical flow, feature matching, and depth estimation, by reducing every two-frame task to predicting a dense 2D displacement field.

desk verdict A strong empirical systems paper that extends the UniMatch-style unification with frozen LVM features and a 1.8M mixed training set; the core claims mostly hold up, but the MegaDepth zero-shot feature-matching result needs a careful train/test scene split check. read the letter →

arxiv 2507.08400 v1 pith:DYIIGUIN submitted 2025-07-11 cs.CV cs.AIcs.MM

Yongjian Zhang , Longguang Wang , Kunhong Li , Ye Zhang , Yun Wang , Liang Lin , Yulan Guo This is my paper
classification cs.CVcs.AIcs.MM
keywords unifiedmatchingmodel2Ddisplacementfieldstereoopticalflowestimationfeaturedepthzero-shotgeneralizationlargevisionmodels
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that a single model with one set of weights can handle all two-frame correspondence tasks—stereo matching, optical flow, feature matching, and depth estimation—on images never seen during training. The enabling move is to discard task-specific geometric priors and treat every matching problem as the prediction of a dense 2D displacement field, from which disparity, depth, and keypoint correspondences are recovered by closed-form formulas. To make an ordinary optical flow architecture generalize, the authors freeze the features of a large vision model and pass them through a transformation pipeline that restores detail and adapts the features to matching. Trained on a reorganized mixture of nearly 1.8 million stereo, flow, and depth samples, the resulting model outperforms the leading unified matching models on cross-task evaluations and matches or beats most task-specific algorithms on their own benchmarks. A reader should care because it suggests a path from many specialized matching networks toward one foundation model deployed across domains.

What carries the argument

The load-bearing mechanism is the unified 2D displacement formulation combined with a feature transformation pipeline that grafts frozen large-vision-model features onto an optical flow architecture. The formulation removes task-specific cost volumes: stereo, flow, feature matching, and depth all become the regression of a dense $(\Delta u, \Delta v)$ field, with parameter-free conversions yielding disparity, keypoint correspondences, and metric depth via a least-squares inversion of the pinhole model. The transformation pipeline preserves the domain-invariance of the frozen encoder: a position-aware guided upsampling block recomposes low-resolution ViT features using attention over nearest neighbours guided by CNN features, a U-shaped fusion adapter with a multi-scale patch embedding block fuses layers into a feature pyramid, and a pixel-wise InfoNCE loss on the cosine-similarity cost volume forces cross-view representations to be consistent. DINOv2-reg-giant supplies the features, FlowFormer supplies the all-pairs cost aggregation, and three training stages progressively move from a single optical flow task to the full 1.8M-sample multi-task mixture and then to fine-grained high-resolution tuning.

What would settle it

Evaluate PanMatch zero-shot on a scene family with no stylistic presence in its training mixture, such as underwater or thermal imagery, with quantitative errors against ground truth; if accuracy falls to the level of the FlowFormer baseline without LVM features, the claimed LVM-driven generalization is not the operative factor. A complementary test trains the identical pipeline on the same 1.8M samples but with a small task-specific encoder: if the gap to PanMatch largely closes, the training mixture, not the LVM, is carrying the result.

Watch

Extended reading notes

Core claim

The paper's central claim is that any two-frame matching task can be restated as all-pairs pixel matching: predict the displacement $(\Delta u, \Delta v)$ from each reference pixel to its target counterpart, and let the task-specific output fall out of that one representation. Disparity is the negative horizontal component along epipolar lines, keypoint correspondences come from applying forward-backward consistency filtering to the dense field, and metric depth is recovered from the displacement plus known camera poses and intrinsics through a least-squares solution of the pinhole projection equations. The authors argue that previous unified attempts failed because they kept task-shaped cost volumes and task-dependent aggregation, so they remove those priors and instead put the burden of generalization on the feature representation. On the evidence they report, the frozen all-purpose features of DINOv2-reg-giant, carried through a guided upsampling block, a feature pyramid adapter, and multi-scale patch embedding, and trained with an additional cross-view contrastive constraint, give an optical flow baseline zero-shot competence on stereo, flow, feature matching, and depth alike: best results on the ETH3D and Spring benchmarks, second place on Middlebury, Sintel, and WxBS against task-specific robust models, and clear wins over UniMatch, CrocoFlow, and Flow-Anything on cross-task tests.

Load-bearing premise

The load-bearing premise is that the frozen large-vision-model features stay domain-invariant after the trainable parts transform them, so the model generalizes rather than memorizing its 1.8-million-sample training data; if that fails, the zero-shot results would be ordinary big-data fine-tuning.

Editorial extensions

If this is right

  • A single PanMatch checkpoint replaces separate stereo, flow, feature-matching, and depth models: the model reports the best zero-shot numbers on the ETH3D and Spring benchmarks among the compared methods, with second-place results on Middlebury, Sintel, and WxBS.
  • Zero-shot robustness extends to scenes where robust-vision-challenge models collapse, including rainy driving footage, night-time streets, satellite imagery, and specular or transparent surfaces, producing usable estimates without any fine-tuning.
  • Cross-task training data is a genuine engine of generalization: on the feature-matching benchmark WxBS the model jumps to 64.2 mAA only once stereo and depth datasets are folded into the mixture, an effect that same-domain flow data alone does not produce.
  • The transformation pipeline is portable: swapping it into PSMNet, RAFT, or FlowFormer in place of their native encoders improves zero-shot stereo and flow generalization by at least 70%, 11%, and 12% respectively under restricted pretraining data.
  • For video and sparse multi-view input, the same weights output per-frame depth without pose information, since reliable correspondences yield relative poses and then metric depth, enabling consistent fused point clouds.

Reading between the lines

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

  • If the frozen large-vision-model features are the true carrier of domain invariance, then scaling the encoder further, such as to larger self-supervised ViTs or encoders pretrained on broader data, should keep lifting zero-shot matching accuracy without any change to the aggregation stack; the paper's own DINOv2-G over DINOv2-B trend is consistent with that prediction.
  • Because the InfoNCE term effectively turns the cost volume into a trained similarity prior, one could test whether the displacement regressor is even necessary for the generalization gain, or whether the transformed features alone give a strong dense matcher.
  • A sharper stress test than anything in the paper would be quantitative evaluation on sensor regimes stylistically absent from the training mixture, such as thermal, underwater, or medical endoscopy imagery; the qualitative rainy, night, and satellite evidence suggests but does not prove invariance there.
  • Because any posed two-view collection can be converted into displacement supervision, the 1.8M-sample mixture is a starting point rather than a ceiling, yet the paper's explanations for the benchmark gaps, that competitors fine-tune on benchmark data, and for the semi-dense depth output, that division in the conversion formulas causes invalid regions, are auxiliary hypotheses that later work shoul
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 / 5 minor

Summary. PanMatch proposes a unified two-frame correspondence model in which stereo matching, optical flow, feature matching, and depth estimation are all cast as 2D displacement prediction with a single set of weights. The architecture uses a frozen DINOv2 encoder whose multi-layer features are passed through a guided upsampling block, an FPN-based fusion adapter, a multi-scale patch embedding, and an InfoNCE cross-view consistency loss, before being fed to a FlowFormer-style aggregation baseline. The model is pretrained on a mixed dataset of nearly 1.8 million samples from optical flow, stereo, and depth sources, with depth and disparity converted to displacement labels via the proposed conversion equations. The paper reports state-of-the-art or competitive zero-shot results on ETH3D, Spring, Middlebury, Sintel, KITTI, ScanNet, MegaDepth, WxBS, and several depth benchmarks, along with ablations supporting the contribution of LVM features and of cross-task training data.

Significance. If the results hold, the paper makes a strong empirical contribution: a single checkpoint that addresses four matching-related tasks with one displacement-based formulation and demonstrates cross-domain generalization beyond task-specific models. The benchmark coverage is extensive, and the ablation in Table 13 is an honest comparison at matched data scales, showing that frozen LVM features help beyond simply adding more data. The parameter-free conversion from displacement to disparity, depth, and keypoint correspondences is elegant and practically useful. The main risks are evaluation-protocol issues that directly affect the strength of the zero-shot claims, particularly the MegaDepth train/test overlap and the depth evaluation protocol.

major comments (3)
  1. [Sec. 5.3.2, Table 9; training data in Table 3] The paper reports feature-matching results on MegaDepth-1500 while Table 3 lists MegaDepth as one of the training sources (~285k samples), but it never states that a scene-level split was enforced between the generated training pairs and the MegaDepth-1500 evaluation pairs. Since MegaDepth-1500 is a subset of image pairs drawn from MegaDepth scenes, training on any pairs from the same scenes would violate the zero-shot premise and inflate the reported AUCs of 50.9/66.8/78.9. This is a load-bearing issue for the abstract and Table 9 claims of zero-shot feature matching on unseen scenes. The authors should specify whether the MegaDepth training samples were split by scene from the MegaDepth-1500 test pairs, and ideally verify non-overlap explicitly; if no scene-level split exists, the corresponding results and claims must be revised.
  2. [Sec. 5.3.3, Table 10] The depth evaluation protocol is not apples-to-apples: the text states that PanMatch is evaluated only on valid regions of each estimated depth map after a depth validity check, because the model may output a semi-dense depth map. The comparison methods in Table 10 are not evaluated under the same validity mask, and no coverage statistics are reported. This makes the claim that PanMatch 'outperforms existing methods in terms of Sq Rel across four datasets' not established. The authors should either report dense full-image metrics, evaluate the baselines under the same valid-region mask, or clearly separate semi-dense results from dense comparisons.
  3. [Sec. 4.2, Fig. 12, Abstract] The abstract claims 'unprecedented zero-shot performance in abnormal scenarios, such as rainy day and satellite imagery,' but the evidence for these extreme domains is qualitative only (Fig. 12). Given that the training mixture includes driving and indoor imagery stylistically close to some of these test scenes, a quantitative evaluation on the rainy, night, satellite, and transparent-object cases is needed to support the zero-shot claim. If quantitative evaluation is not feasible, the claim should be tempered to a qualitative demonstration.
minor comments (5)
  1. [Table 12] The header 'MEP' should be 'MPE' for multi-scale patch embedding.
  2. [Table 3 and Sec. 5.1.2] The caption says datasets used for fine-tuning are highlighted in boldface, but no boldface is visible in the table; please mark them explicitly so the Stage 3 training protocol is unambiguous.
  3. [Sec. 4.2.3, Eq. (14)] The sum over f in the InfoNCE loss is over all proposals, but the domain of f is not explicitly defined; also, 'impress on the other regions' should likely be 'repress' or 'penalize'.
  4. [Sec. 3.2, Eqs. (4)-(8)] The variables T1, T2, R1, R2 are used without an explicit definition of the coordinate frames; please define them (e.g., pose of reference and target cameras in world coordinates) for clarity.
  5. [Fig. 9 caption] The feature-matching metrics in Fig. 9 are not identified; please state which metric (e.g., AUC@5, AUC@10, or mAA@10) is plotted for each dataset.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: PanMatch's central claims are benchmark-tested, and its use of the authors' FormerStereo is transparent, independently published prior work rather than a self-referential load-bearing premise.

full rationale

PanMatch is an empirical systems paper. The load-bearing claim — that a single frozen-LVM model with the proposed feature transformation pipeline generalizes zero-shot across stereo, optical flow, feature matching, and depth — is checked against external benchmarks (Middlebury, ETH3D, KITTI, Sintel, Spring, ScanNet, MegaDepth-1500, WxBS, and depth sets), and no free parameter is fitted to those targets. The 2D displacement unification in Sec. 3 is a definitional reformulation: Eqs. (1)-(8) algebraically convert disparity/depth/flow to displacement and back, so converting training labels and then converting outputs back is not a prediction derived from its own inputs by construction. The feature transformation components (guided feature upsampling, FPN adapter, multi-scale patch embedding, and the InfoNCE loss) are validated by ablations on unseen-domain benchmarks rather than assumed through citation. The paper explicitly states that it extends the authors' ECCV paper FormerStereo [17]; that prior work is independently published and benchmark-tested, and the present contribution lies in extending it to unified matching and evaluating cross-task generalization. This is transparent self-citation, not load-bearing circularity. The reviewer-identified risk that MegaDepth training samples may overlap MegaDepth-1500 test pairs is a data-leakage and correctness concern, not a circularity-by-construction issue, because the evaluation is not equal to the training objective by definition.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The paper introduces no new physical or mathematical entities; it postulates one empirical premise (LVM feature domain invariance after adaptation) and inherits the unified displacement formulation from prior work. The free parameters are design hyperparameters (InfoNCE temperature, contrastive loss weight, confidence threshold, patch sizes) rather than fitted physics constants, with the feature-matching confidence threshold being the most consequential because it determines every feature-matching number reported.

free parameters (4)
  • InfoNCE temperature tau = 0.07
    Temperature in the cross-view contrastive loss (Eq 14), chosen by hand following [40]; its value shapes the proposal distribution and thus the learned feature alignment.
  • Contrastive loss weight = 1.0 (implicit)
    The total loss in Eq 15 is L_disp + L_NCE with no coefficient; the unit weight is a hand choice not justified by ablation.
  • Feature-matching confidence threshold = unspecified
    Displacement outputs are filtered by forward-backward cycle consistency with a threshold (Sec 5.2.2) that is never quantified, yet it determines all feature-matching results in Fig 9 and Table 9.
  • Patch sizes in multi-scale patch embedding = [4, 2, 1, 1/2]
    Hand-chosen patch sizes aligning 1/2, 1/4, 1/8, and 1/16 resolution features into the 1/8-space (Sec 4.2.2); these components are not individually ablated.
assumptions (6)
  • domain assumption All two-frame correspondence tasks can be expressed as dense 2D displacement estimation (Sec 3.1).
    The entire framework rests on the equivalence of stereo disparity, optical flow, and depth to pixel displacements; plausible and inherited from UniMatch [1], but it deliberately discards task-specific constraints (epipolar geometry, rigidity, photometric consistency) that some tasks exploit for accuracy.
  • ad hoc to paper Frozen LVM features are domain-invariant and retain this property after the proposed transformation stack (Sec 4.2).
    The central hypothesis of the paper; supported by ablations (Table 13) on standard benchmarks but only qualitatively on the extreme domains (rain, night, satellite) where it is most load-bearing.
  • domain assumption Camera poses and intrinsics used to convert depth and disparity into displacement labels are accurate (Sec 5.1.1).
    Depth-to-displacement conversion (Eq 4-8) assumes known [R|T] and K, including SfM-derived poses from MegaDepth which contain noise; corrupted labels would bias the learned displacement distribution.
  • domain assumption Forward-backward cycle consistency is a sufficient confidence filter for feature matching (Sec 5.2.2).
    All feature-matching results depend on this filter; the paper does not test alternatives such as learned confidence as in RoMa.
  • domain assumption Benchmark training splits (Middlebury train-h, Sintel training, KITTI training) are valid zero-shot evaluation proxies (Tables 4-5).
    Standard practice in the domain-generalization literature; results on these splits are not official leaderboard scores, so small-number differences should be read cautiously.
  • standard math Pinhole camera projection model for depth back-projection (Eq 4-8).
    Standard projective geometry; the formulas are correct given the stated assumptions of known intrinsics and extrinsics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PanMatch: Unleashing the Potential of Large Vision Models for Unified Matching Models." pith.science (2026). https://pith.science/paper/DYIIGUIN

@misc{pith2026250708400,
  author       = {Pith},
  title        = {Pith review of: PanMatch: Unleashing the Potential of Large Vision Models for Unified Matching Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DYIIGUIN}},
  note         = {Machine review of arXiv:2507.08400}
}
read the original abstract

This work presents PanMatch, a versatile foundation model for robust correspondence matching. Unlike previous methods that rely on task-specific architectures and domain-specific fine-tuning to support tasks like stereo matching, optical flow or feature matching, our key insight is that any two-frame correspondence matching task can be addressed within a 2D displacement estimation framework using the same model weights. Such a formulation eliminates the need for designing specialized unified architectures or task-specific ensemble models. Instead, it achieves multi-task integration by endowing displacement estimation algorithms with unprecedented generalization capabilities. To this end, we highlight the importance of a robust feature extractor applicable across multiple domains and tasks, and propose the feature transformation pipeline that leverage all-purpose features from Large Vision Models to endow matching baselines with zero-shot cross-view matching capabilities. Furthermore, we assemble a cross-domain dataset with near 1.8 million samples from stereo matching, optical flow, and feature matching domains to pretrain PanMatch. We demonstrate the versatility of PanMatch across a wide range of domains and downstream tasks using the same model weights. Our model outperforms UniMatch and Flow-Anything on cross-task evaluations, and achieves comparable performance to most state-of-the-art task-specific algorithms on task-oriented benchmarks. Additionally, PanMatch presents unprecedented zero-shot performance in abnormal scenarios, such as rainy day and satellite imagery, where most existing robust algorithms fail to yield meaningful results.

Figures

Figures reproduced from arXiv: 2507.08400 by the authors.

Figure 1
Figure 1. An overview of the proposed unified architecture. We [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Conversion mechanism. the reference image to the target one. These correspondences are represented using pixel coordinate displacements as: Iref (u1, v1) = Itar(u2 − ∆u, v2 − ∆v), (1) where (∆u, ∆v) denotes the ground-truth displacement vector from the reference image Iref (u1, v1) to the target one Itar(u2, v2). This representation is consistent with the standard definition of optical flow, therefore, the displace￾… view at source ↗
Figure 3
Figure 3. An overview of the proposed framework. ϕ(·) means data augmentation strategies described in Sec. 5.1. 4 UNIFIED MATCHING MODEL With the aforementioned unified formulation, as well as its representational equivalence to optical flow, we develop a unified model based on existing optical flow baselines. However, the limited feature representation capacity of these baselines compromises cross-task and cross-domain gener… view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: Disparity estimation from feature matching without cost aggregation. Features extracted by PSMNet [ [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: An overview of the guided feature upsampling block. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: A pipeline to obtain the patch correspondence distri [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Diversity of the training data. Each row shows paired [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Qualitative results on the ScanNet [92], YFCC [93], and WxBS [94] datasets. All methods employ the same outlier filtering method and confidence threshold. models that support diverse outputs across multiple tasks. Consequently, we select four state-of-the-art optical f…
Figure 9
Figure 9. Figure 9: Zero-shot comparison on multiple domains and tasks, [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]
Figure 10
Figure 10. Figure 10: Comparison on an ETH test scene. Zoom out for [PITH_FULL_IMAGE:figures/full_fig_p009_10.png]
Figure 11
Figure 11. Figure 11: The versatile performance of PanMatch on a wide range of tasks and datasets [ [PITH_FULL_IMAGE:figures/full_fig_p010_11.png]
Figure 12
Figure 12. Figure 12: The real-world zero-shot capacity between RVC algorithms and our proposed PanMatch tested on four abnormal [PITH_FULL_IMAGE:figures/full_fig_p012_12.png]
Figure 14
Figure 14. Figure 14: PCA analysis. Guided Feature Upsampling. In PanMatch, we propose a guided feature upsampling block to recompose feature maps at arbitrary scales under appropriate guidance. This module can be replaced by standard interpolation meth￾ods, such as bilinear interpolation,…
Figure 13
Figure 13. Figure 13: Two-view depth estimation. Each result comes [PITH_FULL_IMAGE:figures/full_fig_p013_13.png]
Figure 15
Figure 15. Figure 15: Generalization enhancement by LNCE. Here, Figs. (a) and (c) are heatmaps generated by feature matching, and (b) and (d) are warped images from another view. encoder, substantially improving both estimation accuracy and generalization capability. These findings align w…
Figure 16
Figure 16. Figure 16: A unique task that the independent methods are [PITH_FULL_IMAGE:figures/full_fig_p015_16.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

126 extracted references · 68 canonical work pages

  1. [1]

    Unifying flow, stereo and depth estimation,

    H. Xu, J. Zhang, J. Cai, H. Rezatofighi, F. Yu, D. Tao, and A. Geiger, “Unifying flow, stereo and depth estimation,”IEEE Trans. Pattern Anal. Mach. Intell., 2023. JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. XX, JUL Y 2025 16

    2023

  2. [2]

    RGM: A Robust Generalizable Matching Model

    S. Zhang, X. Sun, H. Chen, B. Li, and C. Shen, “RGM: A robust generalizable matching model,”arXiv:2310.11755, 2023

    work page Pith review arXiv 2023

  3. [3]

    VGGT: Visual geometry grounded transformer,

    J. Wang, M. Chen, N. Karaev, A. Vedaldi, C. Rupprecht, and D. Novotny, “VGGT: Visual geometry grounded transformer,” inCVPR, 2025

    2025

  4. [4]

    ITSA: An information-theoretic approach to au- tomatic shortcut avoidance and domain generalization in stereo matching networks,

    W. Chuah, R. B. Tennakoon, R. Hoseinnezhad, A. Bab-Hadiashar, and D. Suter, “ITSA: An information-theoretic approach to au- tomatic shortcut avoidance and domain generalization in stereo matching networks,” inCVPR, 2022

    2022

  5. [5]

    Practical stereo matching via cascaded recurrent network with adaptive correlation,

    J. Li, P . Wang, P . Xiong, T. Cai, Z. Yan, L. Yang, J. Liu, H. Fan, and S. Liu, “Practical stereo matching via cascaded recurrent network with adaptive correlation,” inCVPR, 2022

    2022

  6. [6]

    Falling Things: A synthetic dataset for 3D object detection and pose estimation,

    J. Tremblay, T. To, and S. Birchfield, “Falling Things: A synthetic dataset for 3D object detection and pose estimation,” inCVPR Workshops, 2018

    2018

  7. [7]

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

    N. Mayer, E. Ilg, P . H ¨ausser, P . Fischer, D. Cremers, A. Dosovit- skiy, and T. Brox, “A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation,” inCVPR, 2016

    2016

  8. [8]

    FlowNet: Learning optical flow with convolutional networks,

    A. Dosovitskiy, P . Fischer, E. Ilg, P . H ¨ausser, C. Hazirbas, V . Golkov, P . van der Smagt, D. Cremers, and T. Brox, “FlowNet: Learning optical flow with convolutional networks,” inICCV, 2015

    2015

Show all 126 references
  1. [9]

    AutoFlow: Learning a better training set for optical flow,

    D. Sun, D. Vlasic, C. Herrmann, V . Jampani, M. Krainin, H. Chang, R. Zabih, W. T. Freeman, and C. Liu, “AutoFlow: Learning a better training set for optical flow,” inCVPR, 2021

    2021

  2. [10]

    Virtual KITTI 2,

    Y. Cabon, N. Murray, and M. Humenberger, “Virtual KITTI 2,” 2020

    2020

  3. [11]

    Tartanair: A dataset to push the limits of visual SLAM,

    W. Wang, D. Zhu, X. Wang, Y. Hu, Y. Qiu, C. Wang, Y. Hu, A. Kapoor, and S. A. Scherer, “Tartanair: A dataset to push the limits of visual SLAM,” inIROS, 2020

    2020

  4. [12]

    Hypersim: A photore- alistic synthetic dataset for holistic indoor scene understanding,

    M. Roberts, J. Ramapuram, A. Ranjan, A. Kumar, M. ´A. Bautista, N. Paczan, R. Webb, and J. M. Susskind, “Hypersim: A photore- alistic synthetic dataset for holistic indoor scene understanding,” inICCV, 2021

    2021

  5. [13]

    MegaDepth: Learning single-view depth prediction from internet photos,

    Z. Li and N. Snavely, “MegaDepth: Learning single-view depth prediction from internet photos,” inCVPR, 2018

    2018

  6. [14]

    CroCo v2: Improved cross-view completion pre-training for stereo match- ing and optical flow,

    P . Weinzaepfel, T. Lucas, V . Leroy, Y. Cabon, V . Arora, R. Br´egier, G. Csurka, L. Antsfeld, B. Chidlovskii, and J. Revaud, “CroCo v2: Improved cross-view completion pre-training for stereo match- ing and optical flow,” inICCV, 2023

    2023

  7. [15]

    Stereo Anything: Unifying stereo matching with large-scale mixed data,

    X. Guo, C. Zhang, Y. Zhang, D. Nie, R. Wang, W. Zheng, M. Poggi, and L. Chen, “Stereo Anything: Unifying stereo matching with large-scale mixed data,”arXiv:2411.14053, 2024

    2024

  8. [16]

    FoundationStereo: Zero-shot stereo matching,

    B. Wen, M. Trepte, J. Aribido, J. Kautz, O. Gallo, and S. Birchfield, “FoundationStereo: Zero-shot stereo matching,” inCVPR, 2025

    2025

  9. [17]

    Learning representations from foundation models for domain generalized stereo matching,

    Y. Zhang, L. Wang, K. Li, Y. Wang, and Y. Guo, “Learning representations from foundation models for domain generalized stereo matching,” inECCV, 2025

    2025

  10. [18]

    On the synergies between machine learning and binocular stereo for depth estimation from images: A survey,

    M. Poggi, F. Tosi, K. Batsos, P . Mordohai, and S. Mattoccia, “On the synergies between machine learning and binocular stereo for depth estimation from images: A survey,”IEEE Trans. Pattern Anal. Mach. Intell., 2022

    2022

  11. [19]

    A survey on deep learning techniques for stereo-based depth estimation,

    H. Laga, L. V . Jospin, F. Boussa ¨ıd, and M. Bennamoun, “A survey on deep learning techniques for stereo-based depth estimation,” IEEE Trans. Pattern Anal. Mach. Intell., 2022

    2022

  12. [20]

    Computing the stereo matching cost with a convolutional neural network,

    J. Zbontar and Y. LeCun, “Computing the stereo matching cost with a convolutional neural network,” inCVPR, 2015

    2015

  13. [21]

    SGM-Nets: Semi-global matching with neural networks,

    A. Seki and M. Pollefeys, “SGM-Nets: Semi-global matching with neural networks,” inCVPR, 2017

    2017

  14. [22]

    A taxonomy and evaluation of dense two-frame stereo correspondence algorithms,

    D. Scharstein and R. Szeliski, “A taxonomy and evaluation of dense two-frame stereo correspondence algorithms,”Int. J. Com- put. Vis., 2002

    2002

  15. [23]

    Accurate and efficient stereo processing by semi-global matching and mutual information,

    H. Hirschm ¨uller, “Accurate and efficient stereo processing by semi-global matching and mutual information,” inCVPR, 2005

    2005

  16. [24]

    Pyramid stereo matching network,

    J. Chang and Y. Chen, “Pyramid stereo matching network,” in CVPR, 2018

    2018

  17. [25]

    GA- Net: Guided aggregation net for end-to-end stereo matching,

    F. Zhang, V . A. Prisacariu, R. Yang, and P . H. S. Torr, “GA- Net: Guided aggregation net for end-to-end stereo matching,” inCVPR, 2019

    2019

  18. [26]

    AANet: Adaptive aggregation network for efficient stereo matching,

    H. Xu and J. Zhang, “AANet: Adaptive aggregation network for efficient stereo matching,” inCVPR, 2020

    2020

  19. [27]

    Attention concatenation volume for accurate and efficient stereo matching,

    G. Xu, J. Cheng, P . Guo, and X. Yang, “Attention concatenation volume for accurate and efficient stereo matching,” inCVPR, 2022

    2022

  20. [28]

    Stereo matching using multi-level cost volume and multi-scale feature constancy,

    Z. Liang, Y. Guo, Y. Feng, W. Chen, L. Qiao, L. Zhou, J. Zhang, and H. Liu, “Stereo matching using multi-level cost volume and multi-scale feature constancy,”IEEE Trans. Pattern Anal. Mach. Intell., 2021

    2021

  21. [29]

    Hierarchical deep stereo matching on high-resolution images,

    G. Yang, J. Manela, M. Happold, and D. Ramanan, “Hierarchical deep stereo matching on high-resolution images,” inCVPR, 2019

    2019

  22. [30]

    Accurate and efficient stereo matching via attention concatenation volume,

    G. Xu, Y. Wang, J. Cheng, J. Tang, and X. Yang, “Accurate and efficient stereo matching via attention concatenation volume,” IEEE Trans. Pattern Anal. Mach. Intell., 2024

    2024

  23. [31]

    HITNet: Hierarchical iterative tile refinement network for real-time stereo matching,

    V . Tankovich, C. Hane, Y. Zhang, A. Kowdle, S. R. Fanello, and S. Bouaziz, “HITNet: Hierarchical iterative tile refinement network for real-time stereo matching,” inCVPR, 2021

    2021

  24. [32]

    IGEV++: Iterative multi-range geometry encoding volumes for stereo matching,

    G. Xu, X. Wang, Z. Zhang, J. Cheng, C. Liao, and X. Yang, “IGEV++: Iterative multi-range geometry encoding volumes for stereo matching,”IEEE Trans. Pattern Anal. Mach. Intell., 2025

    2025

  25. [33]

    RAFT: recurrent all-pairs field transforms for optical flow,

    Z. Teed and J. Deng, “RAFT: recurrent all-pairs field transforms for optical flow,” inECCV, 2020

    2020

  26. [34]

    RAFT-Stereo: Multilevel recur- rent field transforms for stereo matching,

    L. Lipson, Z. Teed, and J. Deng, “RAFT-Stereo: Multilevel recur- rent field transforms for stereo matching,” in3DV, 2021

    2021

  27. [35]

    Iterative geometry encoding volume for stereo matching,

    G. Xu, X. Wang, X. Ding, and X. Yang, “Iterative geometry encoding volume for stereo matching,” inCVPR, 2023

    2023

  28. [36]

    Selective-Stereo: Adaptive frequency information selection for stereo matching,

    X. Wang, G. Xu, H. Jia, and X. Yang, “Selective-Stereo: Adaptive frequency information selection for stereo matching,” inCVPR, 2024

    2024

  29. [37]

    Domain-invariant stereo matching networks,

    F. Zhang, X. Qi, R. Yang, V . Prisacariu, B. W. Wah, and P . H. S. Torr, “Domain-invariant stereo matching networks,” inECCV, 2020

    2020

  30. [38]

    Matching-space stereo networks for cross-domain generalization,

    C. Cai, M. Poggi, S. Mattoccia, and P . Mordohai, “Matching-space stereo networks for cross-domain generalization,” in3DV, 2020

    2020

  31. [39]

    GraftNet: Towards domain general- ized stereo matching with a broad-spectrum and task-oriented feature,

    B. Liu, H. Yu, and G. Qi, “GraftNet: Towards domain general- ized stereo matching with a broad-spectrum and task-oriented feature,” inCVPR, 2022

    2022

  32. [40]

    Revisiting domain generalized stereo matching networks from a feature consistency perspective,

    J. Zhang, X. Wang, X. Bai, C. Wang, L. Huang, Y. Chen, L. Gu, J. Zhou, T. Harada, and E. R. Hancock, “Revisiting domain generalized stereo matching networks from a feature consistency perspective,” inCVPR, 2022

    2022

  33. [41]

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

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

    2023

  34. [42]

    Domain generalized stereo matching via hierarchical visual transformation,

    T. Chang, X. Yang, T. Zhang, and M. Wang, “Domain generalized stereo matching via hierarchical visual transformation,” inCVPR, 2023

    2023

  35. [43]

    Masked representation learning for domain generalized stereo matching,

    Z. Rao, B. Xiong, M. He, Y. Dai, R. He, Z. Shen, and X. Li, “Masked representation learning for domain generalized stereo matching,” inCVPR, 2023

    2023

  36. [44]

    ProbFlow: Joint optical flow and uncertainty estimation,

    A. S. Wannenwetsch, M. Keuper, and S. Roth, “ProbFlow: Joint optical flow and uncertainty estimation,” inICCV, 2017

    2017

  37. [45]

    PWC-Net: Cnns for optical flow using pyramid, warping, and cost volume,

    D. Sun, X. Yang, M. Liu, and J. Kautz, “PWC-Net: Cnns for optical flow using pyramid, warping, and cost volume,” inCVPR, 2018

    2018

  38. [46]

    Flownet 2.0: Evolution of optical flow estimation with deep networks,

    E. Ilg, N. Mayer, T. Saikia, M. Keuper, A. Dosovitskiy, and T. Brox, “Flownet 2.0: Evolution of optical flow estimation with deep networks,” inCVPR, 2017

    2017

  39. [47]

    Learning to estimate hidden motions with global motion aggregation,

    S. Jiang, D. Campbell, Y. Lu, H. Li, and R. I. Hartley, “Learning to estimate hidden motions with global motion aggregation,” in ICCV, 2021

    2021

  40. [48]

    SEA-RAFT: simple, efficient, accurate RAFT for optical flow,

    Y. Wang, L. Lipson, and J. Deng, “SEA-RAFT: simple, efficient, accurate RAFT for optical flow,” inECCV, 2024

    2024

  41. [49]

    GMFlow: Learning optical flow via global matching,

    H. Xu, J. Zhang, J. Cai, H. Rezatofighi, and D. Tao, “GMFlow: Learning optical flow via global matching,” inCVPR, 2022

    2022

  42. [50]

    FlowFormer: A transformer architecture for optical flow,

    Z. Huang, X. Shi, C. Zhang, Q. Wang, K. C. Cheung, H. Qin, J. Dai, and H. Li, “FlowFormer: A transformer architecture for optical flow,” inECCV, 2022

    2022

  43. [51]

    A naturalistic open source movie for optical flow evaluation,

    D. J. Butler, J. Wulff, G. B. Stanley, and M. J. Black, “A naturalistic open source movie for optical flow evaluation,” inECCV, 2012

    2012

  44. [52]

    Object scene flow,

    M. Menze, C. Heipke, and A. Geiger, “Object scene flow,”ISPRS J. Photogramm. Remote Sens., 2018

    2018

  45. [53]

    Playing for bench- marks,

    S. R. Richter, Z. Hayder, and V . Koltun, “Playing for bench- marks,” inICCV, 2017

    2017

  46. [54]

    Are we ready for au- tonomous driving? The KITTI vision benchmark suite,

    A. Geiger, P . Lenz, and R. Urtasun, “Are we ready for au- tonomous driving? The KITTI vision benchmark suite,” inCVPR, 2012

    2012

  47. [55]

    Kubric: A scalable dataset generator,

    K. Greff, F. Belletti, L. Beyer, C. Doersch, Y. Du, D. Duckworth, D. J. Fleet, D. Gnanapragasam, F. Golemo, C. Herrmann, T. Kipf, A. Kundu, D. Lagun, I. H. Laradji, H. D. Liu, H. Meyer, Y. Miao, D. Nowrouzezahrai, A. C. ¨Oztireli, E. Pot, N. Radwan, D. Rebain, S. Sabour, M. S....

    2022

  48. [56]

    LIFT: learned invariant feature transform,

    K. M. Yi, E. Trulls, V . Lepetit, and P . Fua, “LIFT: learned invariant feature transform,” inECCV, 2016

    2016

  49. [57]

    SuperPoint: Self- supervised interest point detection and description,

    D. DeTone, T. Malisiewicz, and A. Rabinovich, “SuperPoint: Self- supervised interest point detection and description,” inCVPR Workshops, 2018

    2018

  50. [58]

    Super- Glue: Learning feature matching with graph neural networks,

    P . Sarlin, D. DeTone, T. Malisiewicz, and A. Rabinovich, “Super- Glue: Learning feature matching with graph neural networks,” inCVPR, 2020

    2020

  51. [59]

    Decoupling makes weakly supervised local feature better,

    K. Li, L. Wang, L. Liu, Q. Ran, K. Xu, and Y. Guo, “Decoupling makes weakly supervised local feature better,” inCVPR, 2022

    2022

  52. [60]

    Efficient neighbourhood consensus networks via submanifold sparse convolutions,

    I. Rocco, R. Arandjelovi ´c, and J. Sivic, “Efficient neighbourhood consensus networks via submanifold sparse convolutions,” in ECCV, 2020

    2020

  53. [61]

    Dual-resolution corre- spondence networks,

    X. Li, K. Han, S. Li, and V . Prisacariu, “Dual-resolution corre- spondence networks,” inNeurIPS, 2020

    2020

  54. [62]

    LoFTR: Detector- free local feature matching with transformers,

    J. Sun, Z. Shen, Y. Wang, H. Bao, and X. Zhou, “LoFTR: Detector- free local feature matching with transformers,” inCVPR, 2021

    2021

  55. [63]

    RoMa: Robust dense feature matching,

    J. Edstedt, Q. Sun, G. B ¨okman, M. Wadenb¨ack, and M. Felsberg, “RoMa: Robust dense feature matching,” inCVPR, 2024

    2024

  56. [64]

    DKM: dense kernelized feature matching for geometry estima- tion,

    J. Edstedt, I. Athanasiadis, M. Wadenb ¨ack, and M. Felsberg, “DKM: dense kernelized feature matching for geometry estima- tion,” inCVPR, 2023

    2023

  57. [65]

    Semi- dense feature matching with transformers and its applications in multiple-view geometry,

    Z. Shen, J. Sun, Y. Wang, X. He, H. Bao, and X. Zhou, “Semi- dense feature matching with transformers and its applications in multiple-view geometry,”IEEE Trans. Pattern Anal. Mach. Intell., 2023

    2023

  58. [66]

    DI- NOv2: Learning robust visual features without supervision,

    M. Oquab, T. Darcet, T. Moutakanni, H. V . Vo, M. Szafraniec, V . Khalidov, P . Fernandez, D. Haziza, F. Massa, A. El-Nouby, R. Howes, P .-Y. Huang, H. Xu, V . Sharma, S.-W. Li, W. Galuba, M. Rabbat, M. Assran, N. Ballas, G. Synnaeve, I. Misra, H. Je- gou, J. Mairal, P . Labat...

    2023 arXiv

  59. [67]

    Hierarchical discrete distribution decomposition for match density estimation,

    Z. Yin, T. Darrell, and F. Yu, “Hierarchical discrete distribution decomposition for match density estimation,” inCVPR, 2019

    2019

  60. [68]

    Learning accurate dense correspondences and when to trust them,

    P . Truong, M. Danelljan, L. V . Gool, and R. Timofte, “Learning accurate dense correspondences and when to trust them,” in CVPR, 2021

    2021

  61. [69]

    PDC-Net+: Enhanced probabilistic dense correspondence network,

    P . Truong, M. Danelljan, R. Timofte, and L. V . Gool, “PDC-Net+: Enhanced probabilistic dense correspondence network,”IEEE Trans. Pattern Anal. Mach. Intell., 2023

    2023

  62. [70]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agar- wal, G. Sastry, A. Askell, P . Mishkin, J. Clark, G. Krueger, and I. Sutskever, “Learning transferable visual models from natural language supervision,” inICML, 2021

    2021

  63. [71]

    Segment anything,

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y. Loet al., “Segment anything,”arXiv:2304.02643, 2023

    2023 arXiv

  64. [72]

    Depth Anything: Unleashing the power of large-scale unlabeled data,

    L. Yang, B. Kang, Z. Huang, X. Xu, J. Feng, and H. Zhao, “Depth Anything: Unleashing the power of large-scale unlabeled data,” inCVPR, 2024

    2024

  65. [73]

    Vision transformer adapter for dense predictions,

    Z. Chen, Y. Duan, W. Wang, J. He, T. Lu, J. Dai, and Y. Qiao, “Vision transformer adapter for dense predictions,” inICLR, 2023

    2023

  66. [74]

    Convolution meets lora: Parameter efficient finetuning for segment anything model,

    Z. Zhong, Z. Tang, T. He, H. Fang, and C. Yuan, “Convolution meets lora: Parameter efficient finetuning for segment anything model,” inICLR, 2024

    2024

  67. [75]

    Playing to vision foundation model’s strengths in stereo matching,

    C.-W. Liu, Q. Chen, and R. Fan, “Playing to vision foundation model’s strengths in stereo matching,”arXiv:2404.06261, 2024

    2024 arXiv

  68. [76]

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

    L. Bartolomei, F. Tosi, M. Poggi, and S. Mattoccia, “Stereo Any- where: Robust zero-shot deep stereo matching even where either stereo or mono fail,” inCVPR, 2025

    2025

  69. [77]

    Depth Anything V2,

    L. Yang, B. Kang, Z. Huang, Z. Zhao, X. Xu, J. Feng, and H. Zhao, “Depth Anything V2,” inNeurIPS, 2024

    2024

  70. [78]

    SAMFlow: Eliminating any fragmentation in optical flow with segment anything model,

    S. Zhou, R. He, W. Tan, and B. Yan, “SAMFlow: Eliminating any fragmentation in optical flow with segment anything model,” in AAAI, 2024

    2024

  71. [79]

    Decon- volutional networks,

    M. D. Zeiler, D. Krishnan, G. W. Taylor, and R. Fergus, “Decon- volutional networks,” inCVPR, 2010

    2010

  72. [80]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” inNeurIPS, 2017

    2017

  73. [81]

    A ConvNet for the 2020s,

    Z. Liu, H. Mao, C.-Y. Wu, C. Feichtenhofer, T. Darrell, and S. Xie, “A ConvNet for the 2020s,”CVPR, 2022

    2022

  74. [82]

    Momentum contrast for unsupervised visual representation learning,

    K. He, H. Fan, Y. Wu, S. Xie, and R. B. Girshick, “Momentum contrast for unsupervised visual representation learning,” in CVPR, 2020

    2020

  75. [83]

    DynamicStereo: Consistent dynamic depth from stereo videos,

    N. Karaev, I. Rocco, B. Graham, N. Neverova, A. Vedaldi, and C. Rupprecht, “DynamicStereo: Consistent dynamic depth from stereo videos,” inCVPR, 2023

    2023

  76. [84]

    AccFlow: Backward accumulation for long-range optical flow,

    G. Wu, X. Liu, K. Luo, X. Liu, Q. Zheng, S. Liu, X. Jiang, G. Zhai, and W. Wang, “AccFlow: Backward accumulation for long-range optical flow,” inICCV, 2023

    2023

  77. [85]

    Vision trans- formers need registers,

    T. Darcet, M. Oquab, J. Mairal, and P . Bojanowski, “Vision trans- formers need registers,” inICLR, 2024

    2024

  78. [86]

    Decoupled weight decay regulariza- tion,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regulariza- tion,” inICLR, 2019

    2019

  79. [87]

    Flow- Anything: Learning real-world optical flow estimation from large-scale single-view images,

    Y. Liang, Y. Fu, Y. Hu, W. Shao, J. Liu, and D. Zhang, “Flow- Anything: Learning real-world optical flow estimation from large-scale single-view images,”IEEE Trans. Pattern Anal. Mach. Intell., 2025

    2025

  80. [88]

    High-resolution stereo datasets with subpixel-accurate ground truth,

    D. Scharstein, H. Hirschm ¨uller, Y. Kitajima, G. Krathwohl, N. Nesic, X. Wang, and P . Westling, “High-resolution stereo datasets with subpixel-accurate ground truth,” inGCPR, 2014

    2014

  81. [89]

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

    T. Schops, J. L. Schonberger, S. Galliani, T. Sattler, K. Schindler, M. Pollefeys, and A. Geiger, “A multi-view stereo benchmark with high-resolution images and multi-camera videos,” inCVPR, 2017

    2017

  82. [90]

    Infinite photorealistic worlds using procedural generation,

    A. Raistrick, L. Lipson, Z. Ma, L. Mei, M. Wang, Y. Zuo, K. Kayan, H. Wen, B. Han, Y. Wang, A. Newell, H. Law, A. Goyal, K. Yang, and J. Deng, “Infinite photorealistic worlds using procedural generation,” inCVPR, 2023

    2023

  83. [91]

    Spring: A high-resolution high-detail dataset and benchmark for scene flow, optical flow and stereo,

    L. Mehl, J. Schmalfuss, A. Jahedi, Y. Nalivayko, and A. Bruhn, “Spring: A high-resolution high-detail dataset and benchmark for scene flow, optical flow and stereo,” inCVPR, 2023

    2023

  84. [92]

    ScanNet: Richly-annotated 3D reconstructions of indoor scenes,

    A. Dai, A. X. Chang, M. Savva, M. Halber, T. A. Funkhouser, and M. Nießner, “ScanNet: Richly-annotated 3D reconstructions of indoor scenes,” inCVPR, 2017

    2017

  85. [93]

    YFCC100M: the new data in multimedia research,

    B. Thomee, D. A. Shamma, G. Friedland, B. Elizalde, K. Ni, D. Poland, D. Borth, and L. Li, “YFCC100M: the new data in multimedia research,”Commun. ACM, 2016

    2016

  86. [94]

    WxBS: Wide baseline stereo generalizations,

    D. Mishkin, J. Matas, M. Perdoch, and K. Lenc, “WxBS: Wide baseline stereo generalizations,” inBMVC, 2015

    2015

  87. [95]

    Open challenges in deep stereo: the Booster dataset,

    P . Z. Ramirez, F. Tosi, M. Poggi, S. Salti, S. Mattoccia, and L. D. Stefano, “Open challenges in deep stereo: the Booster dataset,” in CVPR, 2022

    2022

  88. [96]

    Stereo corre- spondence and reconstruction of endoscopic data challenge,

    M. Allan, J. Mcleod, C. Wang, J. C. Rosenthal, Z. Hu, N. Gard, P . Eisert, K. X. Fu, T. Zeffiro, W. Xiaet al., “Stereo corre- spondence and reconstruction of endoscopic data challenge,” arXiv:2101.01133, 2021

    2021 arXiv

  89. [97]

    Drivingstereo: A large-scale dataset for stereo matching in au- tonomous driving scenarios,

    G. Yang, X. Song, C. Huang, Z. Deng, J. Shi, and B. Zhou, “Drivingstereo: A large-scale dataset for stereo matching in au- tonomous driving scenarios,” inCVPR, 2019

    2019

  90. [98]

    SUN3D: A database of big spaces reconstructed using sfm and object labels,

    J. Xiao, A. Owens, and A. Torralba, “SUN3D: A database of big spaces reconstructed using sfm and object labels,” inICCV, 2013

    2013

  91. [99]

    A benchmark for the evaluation of RGB-D SLAM systems,

    J. Sturm, N. Engelhard, F. Endres, W. Burgard, and D. Cremers, “A benchmark for the evaluation of RGB-D SLAM systems,” in IROS, 2012

    2012

  92. [100]

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

    Y. Wang, L. Wang, J. Yang, W. An, and Y. Guo, “Flickr1024: A large-scale dataset for stereo image super-resolution,” inCVPR Workshops, 2019

    2019

  93. [101]

    MS- RAFT+: high resolution multi-scale raft,

    A. Jahedi, M. Luz, M. Rivinius, L. Mehl, and A. Bruhn, “MS- RAFT+: high resolution multi-scale raft,”Int. J. Comput. Vis., 2024

    2024

  94. [102]

    Hierarchical object-aware dual-level contrastive learning for domain general- ized stereo matching,

    Y. Miao, M. Wu, S. K. Lam, C. Li, and T. Srikanthan, “Hierarchical object-aware dual-level contrastive learning for domain general- ized stereo matching,” inNeurIPS, 2024

    2024

  95. [103]

    Adaptive multi-modal cross-entropy loss for stereo matching,

    P . Xu, Z. Xiang, C. Qiao, J. Fu, and T. Pu, “Adaptive multi-modal cross-entropy loss for stereo matching,” inCVPR, 2024

    2024

  96. [104]

    FlowFormer++: Masked cost volume autoen- coding for pretraining optical flow estimation,

    X. Shi, Z. Huang, D. Li, M. Zhang, K. C. Cheung, S. See, H. Qin, J. Dai, and H. Li, “FlowFormer++: Masked cost volume autoen- coding for pretraining optical flow estimation,” inCVPR, 2023

    2023

  97. [105]

    SKFlow: Learning optical flow with super kernels,

    S. Sun, Y. Chen, Y. Zhu, G. Guo, and G. Li, “SKFlow: Learning optical flow with super kernels,” inNeurIPS, 2022

    2022

  98. [106]

    CCMR: High res- olution optical flow estimation via coarse-to-fine context-guided motion reasoning,

    A. Jahedi, M. Luz, M. Rivinius, and A. Bruhn, “CCMR: High res- olution optical flow estimation via coarse-to-fine context-guided motion reasoning,” inWACV, 2024

    2024

  99. [107]

    Rethinking optical flow from geometric matching consistent perspective,

    Q. Dong, C. Cao, and Y. Fu, “Rethinking optical flow from geometric matching consistent perspective,” inCVPR, 2023

    2023

  100. [108]

    NLCA-Net: a non-local context attention network for stereo matching,

    Z. Rao, M. He, Y. Dai, Z. Zhu, B. Li, and R. He, “NLCA-Net: a non-local context attention network for stereo matching,”APSIP A Trans. Signal Inf. Process., 2020

    2020

  101. [109]

    CFNet: Cascade and fused cost volume for robust stereo matching,

    Z. Shen, Y. Dai, and Z. Rao, “CFNet: Cascade and fused cost volume for robust stereo matching,” inCVPR, 2021. JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. XX, JUL Y 2025 18

    2021

  102. [110]

    Digging into uncertainty-based pseudo-label for robust stereo matching,

    Z. Shen, X. Song, Y. Dai, D. Zhou, Z. Rao, and L. Zhang, “Digging into uncertainty-based pseudo-label for robust stereo matching,” IEEE Trans. Pattern Anal. Mach. Intell., 2023

    2023

  103. [111]

    An improved raftstereo trained with a mixed dataset for the robust vision challenge 2022,

    H. Jiang, R. Xu, and W. Jiang, “An improved raftstereo trained with a mixed dataset for the robust vision challenge 2022,” arXiv:2210.12785, 2022

    2022 arXiv

  104. [112]

    Uncertainty guided adaptive warping for robust and efficient stereo matching,

    J. Jing, J. Li, P . Xiong, J. Liu, S. Liu, Y. Guo, X. Deng, M. Xu, L. Jiang, and L. Sigal, “Uncertainty guided adaptive warping for robust and efficient stereo matching,” inICCV, 2023

    2023

  105. [113]

    LoS: Local structure-guided stereo matching,

    K. Li, L. Wang, Y. Zhang, K. Xue, S. Zhou, and Y. Guo, “LoS: Local structure-guided stereo matching,” inCVPR, 2024

    2024

  106. [114]

    DEFOM-Stereo: Depth foundation model based stereo match- ing,

    H. Jiang, Z. Lou, L. Ding, R. Xu, M. Tan, W. Jiang, and R. Huang, “DEFOM-Stereo: Depth foundation model based stereo match- ing,”arXiv:2501.09466, 2025

    2025 arXiv

  107. [115]

    Disentangling architecture and training for optical flow,

    D. Sun, C. Herrmann, F. A. Reda, M. Rubinstein, D. J. Fleet, and W. T. Freeman, “Disentangling architecture and training for optical flow,” inECCV, 2022

    2022

  108. [116]

    Booster: A benchmark for depth from images of specular and transparent surfaces,

    P . Z. Ramirez, A. Costanzino, F. Tosi, M. Poggi, S. Salti, S. Mat- toccia, and L. D. Stefano, “Booster: A benchmark for depth from images of specular and transparent surfaces,”IEEE Trans. Pattern Anal. Mach. Intell., 2024

    2024

  109. [117]

    1 year, 1000 km: The oxford robotcar dataset,

    W. Maddern, G. Pascoe, C. Linegar, and P . Newman, “1 year, 1000 km: The oxford robotcar dataset,”The International Journal of Robotics Research, 2017

    2017

  110. [118]

    Semantic stereo for incidental satellite images,

    M. Bosch, K. Foster, G. Christie, S. Wang, G. D. Hager, and M. Brown, “Semantic stereo for incidental satellite images,” in WACV, 2019

    2019

  111. [119]

    ASpanFormer: Detector-free image matching with adaptive span transformer,

    H. Chen, Z. Luo, L. Zhou, Y. Tian, M. Zhen, T. Fang, D. McKin- non, Y. Tsin, and L. Quan, “ASpanFormer: Detector-free image matching with adaptive span transformer,” inECCV, 2022

    2022

  112. [120]

    DeMoN: Depth and motion network for learning monocular stereo,

    B. Ummenhofer, H. Zhou, J. Uhrig, N. Mayer, E. Ilg, A. Doso- vitskiy, and T. Brox, “DeMoN: Depth and motion network for learning monocular stereo,” inCVPR, 2017

    2017

  113. [121]

    DeepMVS: Learning multi-view stereopsis,

    P .-H. Huang, K. Matzen, J. Kopf, N. Ahuja, and J.-B. Huang, “DeepMVS: Learning multi-view stereopsis,” inCVPR, 2018

    2018

  114. [122]

    DPSNet: End-to-end deep plane sweep stereo,

    S. Im, H. Jeon, S. Lin, and I. S. Kweon, “DPSNet: End-to-end deep plane sweep stereo,” inICLR, 2019

    2019

  115. [123]

    Input-level inductive biases for 3D reconstruction,

    W. Yifan, C. Doersch, R. Arandjelovi ´c, J. a. Carreira, and A. Zis- serman, “Input-level inductive biases for 3D reconstruction,” in CVPR, 2022

    2022

  116. [124]

    Swin Transformer: Hierarchical vision transformer using shifted windows,

    Z. Liu, Y. Lin, Y. Cao, H. Hu, Y. Wei, Z. Zhang, S. Lin, and B. Guo, “Swin Transformer: Hierarchical vision transformer using shifted windows,” inICCV, 2021

    2021

  117. [125]

    Twins: Revisiting the design of spatial attention in vision transformers,

    X. Chu, Z. Tian, Y. Wang, B. Zhang, H. Ren, X. Wei, H. Xia, and C. Shen, “Twins: Revisiting the design of spatial attention in vision transformers,” inNeurIPS, 2021

    2021

  118. [126]

    ImageNet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “ImageNet: A large-scale hierarchical image database,” inCVPR, 2009. Yongjian Zhangreceived the B.Eng. degree from Sun Y at-sen University (SYSU) in 2021, where he is currently pursuing the Ph.D. degree with the Sch...

    2009

Pith tools

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