Pith. sign in

REVIEW 1 major objections 6 minor 37 references

Why stereo matching needs both correlation search and feature warping

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 04:28 UTC pith:533XX27W

load-bearing objection A real and useful unification of correlation search and feature warping, with a reproducibility hole: the 'geometry encoding volume' the correlation branch reads from is never defined. the 1 major comments →

arxiv 2607.13674 v1 pith:533XX27W submitted 2026-07-15 cs.CV cs.RO

WAVE-Stereo: Warp-Aligned Volume Encoding for Stereo Matching

Zehan Liu , Yage He , Xianwu Gong This is my paper
classification cs.CV cs.RO
keywords stereo matchingiterative refinementcorrelation volumefeature warpingzero-shot generalizationreal-time inferenceConvGRUglobal context
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

WAVE-Stereo argues that the two dominant correspondence representations in iterative stereo matching—explicit matching search over disparity candidates via correlation volumes, and local residual refinement via warping the right image toward the left—are complementary, and that a single encoder feeding the recurrent update can carry both signals at once. The paper introduces GeoWarp Correspondence Encoder (GWCE) to fuse correlation retrieval, warped-feature alignment, and the current disparity estimate in parallel, and Periodic Global Context Propagation (PGCP) to inject global spatial constraints every few iterations at very low cost. On five unseen real-world benchmarks it reports zero-shot accuracy competitive with foundation-model-based systems at 66 ms inference without any external monocular depth model. The paper also reports that warping alone outperforms correlation alone in its framework, but that fusing both beats either, which is the load-bearing evidence for the central claim.

Core claim

The central claim is that correlation volumes and feature warping are not competing designs but complementary cues, and that a single encoder can feed both to a ConvGRU along with the current disparity estimate. WAVE-Stereo reports this unified encoding, called GWCE, reduces SceneFlow EPE to 0.43 and yields zero-shot KITTI-15 D1-all of 3.18% and Booster Bad-2.0 of 4.42% at 66 ms, without an external foundation model. The paper further claims that periodically adding global self-attention and a lightweight DPT-style fusion to the hidden state every three iterations improves textureless-region accuracy at negligible overhead.

What carries the argument

GeoWarp Correspondence Encoder (GWCE): three parallel branches at the ConvGRU input—a correlation branch retrieving local matching costs, a disparity-prior branch encoding the current estimate with a large-kernel convolution, and a cross-view warping branch that warps right-image features to the left using the current disparity and encodes their residual—fused into motion features. Periodic Global Context Propagation (PGCP): every K iterations, motion features are pooled to 1/32 resolution, processed by a compact self-attention Transformer, fused by a simplified DPT module, and gated into the GRU hidden state via a learnable scalar. The paper does not define the construction of the 'geometry

Load-bearing premise

The correlation branch of GWCE assumes a defined 'geometry encoding volume' from which matching costs are retrieved, but the paper never specifies how that volume is constructed, so the contribution of the correlation branch cannot be reproduced or fully attributed.

What would settle it

Train WAVE-Stereo with the geometry encoding volume replaced by the raw all-pairs correlation volume and with the GWCE warping branch removed. If SceneFlow EPE and KITTI-15 D1-all do not move, the complementarity claim is not supported; conversely, if the warping branch alone matches the full model, the correlation branch is not the source of the gain.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If correct, real-time zero-shot stereo matching can be achieved without a monocular depth foundation model, making edge-deployment feasible.
  • Unifying matching-search and residual-alignment signals in one encoder is a reusable recipe for other iterative correspondence tasks, such as optical flow.
  • Periodic low-resolution global injection is a general way to give recurrent estimators a global receptive field at near-constant cost.
  • The warping branch's reported edge over the correlation branch suggests residual alignment cues deserve more weight in iterative stereo design.
  • Single-stage supervised training on diverse synthetic data can match multi-stage distillation approaches in zero-shot generalization.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The reported accuracy depends on a 'geometry encoding volume' whose construction is never specified; if it silently falls back to the raw all-pairs correlation, the claimed retrieval of geometric priors would need to be re-scoped.
  • The periodic-injection design suggests a testable extension: applying PGCP at multiple resolutions or to the context features, not just the hidden state, could further help in large textureless regions.
  • If the warping branch's per-iteration cost scales with resolution, the 66 ms figure at 720p may not transfer to higher-resolution inputs, where the disparity limit of 192 already restricts application.
  • The complementarity claim implies that methods built purely on either representation leave a consistent accuracy margin on ambiguous surfaces; that margin is measurable by re-running the branch ablation on Booster or Middlebury rather than SceneFlow.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

1 major / 6 minor

Summary. WAVE-Stereo proposes a lightweight iterative stereo matching architecture that combines three cues at the input of a ConvGRU: correlation features retrieved from a geometry encoding volume, the current disparity estimate, and warped cross-view features. It also introduces Periodic Global Context Propagation (PGCP), which injects global self-attention features at 1/32 resolution every K iterations. The authors evaluate zero-shot generalization on Middlebury, ETH3D, KITTI 2012/2015, and Booster, reporting a competitive accuracy-efficiency trade-off (e.g., 3.18% D1-all on KITTI 2015 at 66 ms) without using an external monocular depth foundation model. Ablations on SceneFlow are used to support the contributions of GWCE and PGCP.

Significance. The central claim—that correlation-volume search and feature-warping residual alignment are complementary and can be fused in a lightweight recurrent stereo architecture—is interesting and timely. If the architecture is fully specified, the paper would offer a useful data point: competitive zero-shot accuracy without 3D cost-volume aggregation, monocular foundation models, or multi-stage distillation. The strengths of the paper are its extensive zero-shot evaluation on five real-world benchmarks, the runtime decomposition, and the component ablations (Tables 3–5). The reported speed of 66 ms on an RTX 3090 with no external prior is a practically meaningful result. The code is promised, which would aid reproducibility. However, the paper's current form omits the construction of the geometry encoding volume used by the correlation branch, which is a load-bearing component; until this is specified, the contribution of GWCE and the reported architecture cannot be fully assessed.

major comments (1)
  1. [§3.1 (Eq. 1) and §3.3] The correlation branch is defined as retrieving 'local correlation features c_t' from 'the geometry encoding volume,' and the iterative update (Sec. 3.3) repeats this phrase. However, the construction of this volume is never given. Section 3.3 only describes computing an all-pairs cost volume C and regularizing it with a 'lightweight 2D aggregation module' before soft-argmin regression; no multi-scale geometry encoding volume, no 3D hourglass aggregation, and no retrieval/indexing operation (e.g., sampling around d_{t-1}) are specified. Related Work (Sec. 2.1) states the method 'uses the geometry encoding volume proposed by IGEV,' but the Method section does not say whether IGEV's volume is adopted unchanged, whether C itself serves as the volume, or how the branch's input is constructed. This is not a cosmetic omission: Table 4 attributes a 3.8% EPE improvement to the correlation branch
minor comments (6)
  1. [§4.2] The training list is described as '9 public synthetic datasets,' yet InStereo2K [8] is a real-world indoor stereo dataset. Please correct the terminology and adjust the Limitations statement accordingly.
  2. [References] References [1]–[5] appear as 'Anonymous' instead of author names. This prevents verification of the related-work claims and should be fixed.
  3. [Eq. (7)] The notation softmax(C(d)) is ambiguous; softmax should be over the disparity dimension for each spatial location.
  4. [§3.3] The 'lightweight 2D aggregation module' is described only as following LightStereo [12]. For reproducibility, give the channel counts, number of blocks, and any modifications.
  5. [Table 1 note] It would be helpful to state the inference iteration count for each iterative baseline, not just WAVE-Stereo, given the reported speed comparison.
  6. [Abstract/Title] The abstract and title contain typographical spacing artifacts ('W A VE-Stereo', 'GeoWarp ...') that should be cleaned up in the camera-ready.

Circularity Check

0 steps flagged

No significant circularity: empirical architecture study validated on external benchmarks; the undefined geometry encoding volume is a reproducibility gap, not a circular step.

full rationale

The paper's central claims—that GWCE unifies correlation retrieval and feature warping, and that PGCP improves textureless-region accuracy—are supported by external benchmarks (Table 1: KITTI, Middlebury, ETH3D, Booster) and by internal ablations (Tables 3–5) on SceneFlow. These evaluations are not constructed from the method's own outputs: the zero-shot numbers are measured on unseen real-world datasets, and the ablations compare architecture variants on a held-out synthetic test set. Hyperparameters such as K=3 and Transformer hidden dimension 128 are chosen via SceneFlow ablation, but they do not encode the zero-shot results. The paper's use of IGEV's geometry encoding volume is a citation to external prior work ('WAVE-Stereo follows the ConvGRU iterative optimization paradigm pioneered by RAFT-Stereo [15] and uses the geometry encoding volume proposed by IGEV [32] for matching cost retrieval'), not a self-citation chain, and no derivation step reduces to its own input by construction. The most serious flaw is that Section 3.1/Section 3.3 invoke 'the geometry encoding volume' without specifying how it is built from the 3D cost volume C, making the architecture hard to reproduce exactly; however, this is a completeness/reproducibility limitation, not circular reasoning. The anonymous self-citations (e.g., WAFT-Stereo [5]) are used as related-work comparisons, not as the justificatory basis for the paper's own claims. Therefore no circular step can be exhibited, and the appropriate score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central claim relies on a small set of hyperparameters tuned on the internal SceneFlow ablation, and on domain assumptions about the portability of the geometry encoding volume and the sufficiency of the training data. No non-standard mathematical axioms or invented physical entities are introduced. The free parameters are architectural choices selected by validation, not fitted constants that directly encode the benchmark results.

free parameters (4)
  • PGCP injection interval K = 3
    Chosen via ablation (Table 5) on SceneFlow test set; K=3 gives best EPE among tested values.
  • PGCP Transformer hidden dimension = 128
    Selected via ablation (Table 5); dim=128 yields best EPE 0.43.
  • PGCP Transformer layers/heads/MLP ratio = 3 layers, 4 heads, ratio 3
    Configuration b in Table 5, chosen by ablation to optimize SceneFlow EPE.
  • Loss weights lambda_0 and gamma = 0.5, 0.9
    Set manually (Eq. 8); these hyperparameters are not motivated by analysis.
axioms (4)
  • domain assumption A 'geometry encoding volume' as in IGEV-Stereo exists and provides informative correlation features for the GWCE branch.
    The method references this volume (Section 3.1, Eq. 1) but never defines its construction, relying on prior work's design being ported or available.
  • domain assumption Training on nine synthetic datasets (SceneFlow, FSD, FallingThings, etc.) is sufficient for zero-shot generalization to the five target benchmarks.
    The paper states this in Section 4.2 and claims zero-shot success, but does not provide evidence that other training data choices would not change the outcome.
  • domain assumption The evaluation protocols (Middlebury half-res, ETH3D non-occluded, Booster quarter-res) are standard and comparable across methods.
    The paper states these in Section 4.3 and Table 1 notes, but does not justify the choice over other evaluation settings.
  • standard math Bilinear warping, soft-argmin, and ConvGRU operations are differentiable and function as standard in iterative stereo matching.
    These are standard building blocks from RAFT-Stereo and IGEV-Stereo, used without proof.

pith-pipeline@v1.3.0-alltime-deepseek · 12855 in / 9924 out tokens · 622030 ms · 2026-08-02T04:28:16.452941+00:00 · methodology

0 comments
read the original abstract

Existing iterative stereo matching methods primarily adopt two types of correspondence representation: explicit matching search via correlation volumes and local residual refinement via warped features, yet the two remain separately modeled. We propose WAVE-Stereo, built on a core insight: correlation volumes and feature warping provide complementary matching cues. \textbf{GeoWarp Correspondence Encoder (GWCE)} encodes matching search, residual alignment, and disparity prior in parallel at the ConvGRU input. To mitigate matching degradation in textureless regions, we propose \textbf{Periodic Global Context Propagation (PGCP)}, which propagates global spatial information in a periodic manner. On five real-world benchmarks -- Middlebury, ETH3D, KITTI 2012, KITTI 2015, and Booster -- WAVE-Stereo achieves competitive zero-shot generalization accuracy without any external foundation model prior, achieving 3.18\% D1-all on KITTI 2015, 4.42\% Bad-2.0 on Booster, and 66ms real-time inference, striking a favorable balance between accuracy and efficiency. Our code is available at https://github.com/yamanoko-do/WAVE-Stereo.

Figures

Figures reproduced from arXiv: 2607.13674 by Xianwu Gong, Yage He, Zehan Liu.

Figure 1
Figure 1. Figure 1: Speed-accuracy trade-off on KITTI 2015. WAVE [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Zero-shot evaluation on five real-world benchmarks. Our [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overall architecture of WAVE-Stereo. The pipeline consists of two key components: GWCE unifies correlation search, residual [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visualized qualitative comparisons for real-time disparity estimation methods are displayed on Middlebury, ETH3D, Booster and [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Detailed architecture of the GeoWarp Correspondence [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Ablation visualization of the PGCP module. From left [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

37 extracted references · 1 linked inside Pith

  1. [1]

    Bridgedepth: Unifying monocular and stereo depth estimation

    Anonymous. Bridgedepth: Unifying monocular and stereo depth estimation. InIEEE Conf. Comput. Vis. Pattern Recog., 2025. 5

    2025

  2. [2]

    Waft: Warping-alone field transforms for opti- cal flow

    Anonymous. Waft: Warping-alone field transforms for opti- cal flow. InIEEE Conf. Comput. Vis. Pattern Recog., 2025. 2, 3

    2025

  3. [3]

    Fast-foundationstereo: Real-time zero-shot stereo matching

    Anonymous. Fast-foundationstereo: Real-time zero-shot stereo matching. InIEEE Conf. Comput. Vis. Pattern Recog.,

  4. [4]

    Lite any stereo: Efficient zero-shot stereo matching

    Anonymous. Lite any stereo: Efficient zero-shot stereo matching. InIEEE Conf. Comput. Vis. Pattern Recog., 2026. 4, 5, 9

    2026

  5. [5]

    Waft-stereo: Warping-alone field transforms for stereo matching

    Anonymous. Waft-stereo: Warping-alone field transforms for stereo matching. InIEEE Conf. Comput. Vis. Pattern Recog., 2026. 2, 3, 6, 8, 9

    2026

  6. [6]

    Cleardepth: En- hanced stereo perception of transparent objects for robotic manipulation.arXiv preprint arXiv:2409.08926, 2024

    Kaixin Bai, Huajian Zeng, Lei Zhang, Yiwen Liu, Hongli Xu, Zhaopeng Chen, and Jianwei Zhang. Cleardepth: En- hanced stereo perception of transparent objects for robotic manipulation.arXiv preprint arXiv:2409.08926, 2024. 8

    arXiv 2024

  7. [7]

    Correlate-and- excite: Real-time stereo matching via guided cost volume excitation

    Antyanta Bangunharcana, Jae Won Cho, Seokju Lee, In So Kweon, Kyung-Soo Kim, and Soohyun Kim. Correlate-and- excite: Real-time stereo matching via guided cost volume excitation. In2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3542–3548,

  8. [8]

    In- stereo2k: A large real dataset for stereo matching in indoor scenes.SCIS, 2020

    Wenjing Bao, Wei Xu, Xiaoxue Chen, and Yulan Guo. In- stereo2k: A large real dataset for stereo matching in indoor scenes.SCIS, 2020. 8

    2020

  9. [9]

    Vir- tual kitti 2.arXiv, 2020

    Yohann Cabon, Naila Murray, and Martin Humenberger. Vir- tual kitti 2.arXiv, 2020. 8

    2020

  10. [10]

    Monster++: A unified geometric foun- dation model for multi-view depth estimation via harness- ing monodepth priors

    Junda Cheng et al. Monster++: A unified geometric foun- dation model for multi-view depth estimation via harness- ing monodepth priors. InIEEE Conf. Comput. Vis. Pattern Recog., 2025. 5, 8, 9

    2025

  11. [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. InIEEE Conf. Comput. Vis. Pattern Recog., 2012. 8

    2012

  12. [12]

    Light- stereo: Channel boost is all you need for efficient stereo matching

    Xiuyan Guo, Chenming Zhang, Dujun Nie, Wenzhao Zheng, Yuelin Li, Xu Liu, Yuyao Lin, Feiyue Zhou, et al. Light- stereo: Channel boost is all you need for efficient stereo matching. InIEEE Conf. Comput. Vis. Pattern Recog., 2025. 4, 5, 8, 9

    2025

  13. [13]

    Stereonet: Guided hierarchical refinement for real-time edge-aware depth prediction

    Sameh Khamis, Sean Fanello, Christoph Rhemann, Adarsh Kowdle, Julien Valentin, and Shahram Izadi. Stereonet: Guided hierarchical refinement for real-time edge-aware depth prediction. InEur. Conf. Comput. Vis., 2018. 4

    2018

  14. [14]

    Practical stereo matching via cascaded recurrent net- work with adaptive correlation

    Jiankun Li, Peisen Wang, Pengfei Xiong, Tao Cai, Ziwei Yan, Lei Yang, Jiangyu Liu, Haoqiang Fan, and Shuaicheng Liu. Practical stereo matching via cascaded recurrent net- work with adaptive correlation. InIEEE Conf. Comput. Vis. Pattern Recog., 2022. 8

    2022

  15. [15]

    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 3DV, 2021. 1, 3, 6, 8, 9

    2021

  16. [16]

    Decoupled weight decay regularization.Int

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization.Int. Conf. Learn. Represent., 2019. 8

    2019

  17. [17]

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

    Nikolaus Mayer, Eddy Ilg, Philip H ¨ausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. InIEEE Conf. Com- put. Vis. Pattern Recog., 2016. 7, 8

    2016

  18. [18]

    Object scene flow for autonomous vehicles

    Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. InIEEE Conf. Comput. Vis. Pattern Recog., 2015. 8

    2015

  19. [19]

    Booster: A benchmark for depth from images of specular and transpar- ent surfaces.IEEE Trans

    Pierluigi Zama Ramirez, Fabio Tosi, Matteo Poggi, Samuele Salti, Luigi Di Stefano, and Stefano Mattoccia. Booster: A benchmark for depth from images of specular and transpar- ent surfaces.IEEE Trans. Pattern Anal. Mach. Intell., 2023. 8

    2023

  20. [20]

    Mobilenetv2: Inverted residuals and linear bottlenecks

    Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh- moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. InIEEE Conf. Comput. Vis. Pattern Recog., 2018. 4, 5, 7

    2018

  21. [21]

    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. InGCPR, 2014. 8

    2014

  22. [22]

    Sch¨onberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and An- dreas Geiger

    Thomas Sch ¨ops, Johannes L. Sch¨onberger, 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. InIEEE Conf. Comput. Vis. Pattern Recog., 2017. 8

    2017

  23. [23]

    Mobilestereonet: Towards lightweight deep net- works for stereo matching

    Faranak Shamsafar, Samuel Woerz, Rafia Rahim, and An- dreas Zell. Mobilestereonet: Towards lightweight deep net- works for stereo matching. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2417–2426, 2022. 4

    2022

  24. [24]

    Raft: Recurrent all-pairs field transforms for optical flow

    Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. InEur. Conf. Comput. Vis., 2020. 3, 7

    2020

  25. [25]

    Smd-nets: Stereo mixture density networks

    Fabio Tosi, Yiyi Liao, Carolin Schmitt, and Andreas Geiger. Smd-nets: Stereo mixture density networks. InConference on Computer Vision and Pattern Recognition (CVPR), 2021. 8

    2021

  26. [26]

    Falling things: A synthetic dataset for 3d object detection and pose estimation

    Jonathan Tremblay, Thang To, and Stan Birchfield. Falling things: A synthetic dataset for 3d object detection and pose estimation. InEur. Conf. Comput. Vis., 2018. 8

    2018

  27. [27]

    IRS: A large naturalistic indoor robotics stereo dataset to train deep models for dis- parity and surface normal estimation

    Qiang Wang, Shizhen Zheng, Qingsong Yan, Fei Deng, Kaiyong Zhao, and Xiaowen Chu. IRS: A large naturalistic indoor robotics stereo dataset to train deep models for dis- parity and surface normal estimation. InIEEE International Conference on Multimedia and Expo (ICME), 2021. 8

    2021

  28. [28]

    Tartanair: A dataset to push the limits of visual slam

    Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuqiu Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Se- bastian Scherer. Tartanair: A dataset to push the limits of visual slam. InIROS, 2020. 8

    2020

  29. [29]

    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. InIEEE Conf. Comput. Vis. Pattern Recog., 2024. 8, 9

    2024

  30. [30]

    Foundationstereo: Zero-shot stereo matching

    Bowen Wen, Alex Trevithick, Jan Kautz, Stan Birchfield, and Jonathan Tremblay. Foundationstereo: Zero-shot stereo matching. InIEEE Conf. Comput. Vis. Pattern Recog., 2025. 5, 8, 9

    2025

  31. [31]

    Bilateral grid learning for stereo matching networks

    Bin Xu, Yuhua Xu, Xiaoli Yang, Wei Jia, and Yulan Guo. Bilateral grid learning for stereo matching networks. InIEEE Conf. Comput. Vis. Pattern Recog., 2021. 4

    2021

  32. [32]

    Iterative geometry encoding volume for stereo matching

    Gangwei Xu, Xianqi Wang, Xiaohuan Ding, and Xin Yang. Iterative geometry encoding volume for stereo matching. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 2, 3, 6, 8, 9

    2023

  33. [33]

    Accurate and efficient stereo matching via attention concatenation volume.IEEE Transactions on Pattern Anal- ysis and Machine Intelligence, 2023

    Gangwei Xu, Yun Wang, Junda Cheng, Jinhui Tang, and Xin Yang. Accurate and efficient stereo matching via attention concatenation volume.IEEE Transactions on Pattern Anal- ysis and Machine Intelligence, 2023. 4

    2023

  34. [34]

    Igev++: Iterative multi-range geometry en- coding volumes for stereo matching.IEEE Trans

    Gangwei Xu, Xianqi Wang, Zhaojie Zhang, Xiaohuan Ding, and Xin Yang. Igev++: Iterative multi-range geometry en- coding volumes for stereo matching.IEEE Trans. Pattern Anal. Mach. Intell., 2025. 3, 8, 9

    2025

  35. [35]

    Aanet: Adaptive aggrega- tion network for efficient stereo matching

    Haofei Xu and Juyong Zhang. Aanet: Adaptive aggrega- tion network for efficient stereo matching. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1959–1968, 2020. 4

    1959

  36. [36]

    Matchattention: Matching the relative positions for high-resolution cross-view matching.arXiv preprint arXiv:2510.14260, 2025

    Tingman Yan, Tao Liu, Xilian Yang, Qunfei Zhao, and Zeyang Xia. Matchattention: Matching the relative positions for high-resolution cross-view matching.arXiv preprint arXiv:2510.14260, 2025. 9

    Pith/arXiv arXiv 2025

  37. [37]

    Depth any- thing v2

    Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiao- gang Xu, Jiashi Feng, and Hengshuang Zhao. Depth any- thing v2. InAdv. Neural Inform. Process. Syst., 2024. 2, 5, 8

    2024