Pith. sign in

REVIEW 3 major objections 5 minor 3 cited by

RobuSTereo: Robust Zero-Shot Stereo Matching under Adverse Weather

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 RobuSTereo claims that training stereo matchers on diffusion-generated adverse-weather pairs with a denoising transformer encoder yields state-of-the-art zero-shot robustness to rain, fog, and snow.

desk verdict Plausible and useful contribution, but the missing geometric validation of generated pairs is a real gap and the 'unlimited data' claim is overblown. read the letter →

arxiv 2507.01653 v1 pith:NSPGQLFW submitted 2025-07-02 cs.CV

Yuran Wang , Yingping Liang , Yutao Hu , Ying Fu This is my paper
classification cs.CV
keywords stereomatchingzero-shotgeneralizationadverseweatherdiffusion-baseddatagenerationconsistencyrobustfeatureencoderdisparityestimationdomaingap
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 two obstacles to zero-shot stereo matching in adverse weather—scarce labeled training data and feature extractors that fail on degraded images—can both be addressed by one framework. It generates synthetic rainy, foggy, and snowy stereo pairs from clean pairs using a text-conditioned diffusion model guided by monocular depth, and adds a patch-fusion step that keeps the left and right views geometrically consistent. It then trains standard stereo networks on this synthetic data and replaces their encoder with a hybrid that combines VGG19 local features with a denoising vision transformer, which suppresses weather-induced noise at the feature level. In the paper's experiments the trained model reports EPE 0.836 and D1 1.598 on DrivingStereo overall, outperforming strong zero-shot baselines including StereoAnything, and improves results on the SeeingThroughFog benchmark as well. If these results hold, the framework would make existing stereo matchers weather-robust without collecting new real-world adverse-weather data.

What carries the argument

The method's carrying mechanism is a two-part design. On the data side, a depth-conditioned latent diffusion generator (Stable Diffusion with ControlNet, conditioned on depth from DepthAnythingV2) re-synthesizes each clean stereo view under weather prompts, and a 'Disparity Fusion Method' (DFM) matches patches between left and right, fuses the top-n most similar ones, passes them through self-attention, then un-fuses them; this forces cross-view consistency so the original disparity map remains a usable training label. On the model side, a robust feature encoder combines VGG19's multi-scale local features with a denoising vision transformer (DVT), which reduces noise at the feature level before the disparity refinement network produces the final estimate.

What would settle it

Compute the epipolar error or forward-backward consistency of the generated pairs (I'_L, I'_R) against the original disparity map D: if a non-trivial fraction of pixels violate the epipolar constraint by more than a few pixels, the supervised signal from D is corrupted and the reported zero-shot gains would not be expected to transfer to real degraded pairs with correct geometry.

Watch

Extended reading notes

Core claim

The central claim is that training stereo networks on diffusion-generated adverse-weather pairs, together with a robust encoder, gives state-of-the-art zero-shot performance in unseen weather. The authors present quantitative results on DrivingStereo and SeeingThroughFog where their model, trained on the RST-Dataset, beats strong baselines on rainy, foggy, cloudy, and snowy subsets. The paper's own ablations attribute the gain to two components: the stereo consistency module, which when removed raises EPE from 0.875 to 1.308 for StereoBase, and the robust encoder, which lowers overall EPE from 0.875 (MobileNetV2) and 0.852 (DinoV2) to 0.836. The discovery is an application-level one: a controllable diffusion pipeline can produce training data that transfers to real weather better than physics-based simulation or real LiDAR-labeled data, provided the generated left–right pairs stay geometrically aligned.

Load-bearing premise

The approach stands on the assumption that the diffusion-generated left and right views still share the original scene geometry, so the original disparity map remains a valid training label.

Editorial extensions

If this is right

  • If the central claim holds, any existing stereo network can be made weather-robust by fine-tuning on the generated RST-Dataset, avoiding the cost and label-noise of real adverse-weather collection campaigns.
  • The paper's ablation makes the consistency module load-bearing: without it, EPE on DrivingStereo degrades from 0.875 to 1.308 for StereoBase, indicating that geometric alignment of generated pairs is necessary for the training signal to be valid.
  • The robust encoder is a separate source of gain: in identical training settings it improves overall EPE from 0.875 (MobileNetV2) and 0.852 (DinoV2) to 0.836, suggesting feature-level denoising matters even when the training data is already weather-degraded.
  • Dataset source matters for generation quality: generated pairs derived from real KITTI images outperform generated pairs from the synthetic vKITTI, implying texture realism in the source images carries into the synthesized weather data.

Reading between the lines

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

  • The paper leaves implicit that its gains depend on the generated left and right views actually sharing the original scene geometry; a direct epipolar-error measurement on generated pairs would separate appearance transfer from true geometric preservation.
  • Since the generation is prompt-driven, the same pipeline should transfer to other degradations such as nighttime, underwater, or motion blur, with the consistency module doing the same geometric enforcement.
  • A cheaper alternative worth testing is whether depth-conditioning alone, without the patch-fusion module, would suffice if a post-hoc stereo rectification or optical-flow warp were applied to generated pairs, which would isolate the incremental value of DFM.
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. The paper proposes RobuSTereo, a framework to improve zero-shot stereo matching under adverse weather. It has two main contributions: (i) a diffusion-based data generation pipeline (Stable Diffusion + ControlNet) that turns clean stereo pairs into adverse-weather pairs under monocular-depth guidance, augmented by a "coherence-enhanced consistency module" that fuses top-n similar left/right patches; and (ii) a robust feature encoder combining VGG19 with a Denoising Vision Transformer (DVT) for feature extraction under degraded conditions. The authors generate an RST-Dataset from KITTI and vKITTI, train stereo models on it, and evaluate zero-shot on DrivingStereo and SeeingThroughFog. They report large improvements, e.g., EPE 0.836 / D1 1.598 on DrivingStereo versus EPE 1.042 / D1 3.865 for StereoAnything.

Significance. If the reported results are reliable, the paper would make a useful contribution to stereo matching in adverse weather: the synthetic data pipeline could be reused beyond this paper, and the encoder design is simple to integrate. The experimental evaluation is broad, including comparisons across many recent stereo networks and two real adverse-weather datasets, and the ablation study covers the consistency module, data source, and encoder. The main limitations are the lack of any verification that the generated stereo pairs preserve the source disparity ground truth, the absence of controlled comparisons that vary training data and architecture separately, and the underspecification of the core consistency module. No code or generated dataset is released, which limits reproducibility.

major comments (3)
  1. [Section 3.2, Eqs. (1)-(7)] The load-bearing premise is that the generated pairs (I'_L, I'_R) form a valid rectified stereo pair with the same disparity field as the source ground truth D. However, the left and right views are generated largely independently, each conditioned on monocular depth predictions from DepthAnythingV2 (Eq. 1), and monocular depth is scale-ambiguous and estimated per view. The consistency module (Eqs. 4-7) fuses only the top-n similar patches through self-attention, which can increase visual similarity without enforcing epipolar geometry or preserving the original disparity. No quantitative evaluation of the generated pairs is provided: no epipolar error, no vertical-disparity histogram, no comparison between DGT and the actual correspondence in the generated pair. Since training uses L1 loss against the original DGT, any geometric deviation corrupts the supervision. Please add a quantitative fidelity check (e.g., run a pretrained stereo matcher on the generated pairs and compare its output to DGT, or measure vertical disparity/epipolar error on a sample of generated pairs) and report the statistics. If the deviations are large, the source of the reported gains needs to be reconsidered.
  2. [Section 4.1 and Table 1] The headline comparison in Table 1 is confounded: the proposed model differs from most baselines in both training data and architecture. The caption says most methods are trained on SceneFlow, while "Ours" is trained on the generated dataset; Section 4.1 states "All models are trained on the SceneFlow dataset and our RST-dataset," which is ambiguous. A reader cannot tell whether the improvement comes from the RST-Dataset, the robust encoder, or both. Table 3 partially controls for training data by fixing the network, and Table 4 abates the encoder, but the two are never combined in a single controlled experiment. Please provide a controlled comparison in which the same base network is trained on (i) SceneFlow only, (ii) SceneFlow + a simple weather-augmentation baseline (e.g., synthetic rain/fog overlays), and (iii) SceneFlow + RST-Dataset, and in which the proposed encoder is also evaluated with the original training sets. This would isolate the contribution of each component.
  3. [Section 3.2, Eqs. (4)-(7)] The consistency module is not specified sufficiently for reproduction. The matching in Eq. (4) is described as based on "disparity similarity and image similarity," but no formula, patch size, stride, or number of patches is given; the value of n (top-n) is never stated; and the fusion/un-fusion operations M(·) and U(·) are not defined formally. Since this module is a core contribution claimed to preserve stereo consistency, the paper should either provide complete equations and hyperparameters or release the code and generation pipeline. Without this, the data-generation method cannot be evaluated or built upon.
minor comments (5)
  1. [Section 4.1] The composition of RST-Dataset is not specified: how many source images were used, how many generated pairs were produced per source image, what image resolutions were used, and what text prompts were collected beyond the rainy example in Section 3.2.
  2. [Table 4] The layout of Table 4 is confusing: the rows for "Consistency Module," "Source Data," and "Network Encoder" do not make clear which components are fixed in each comparison, and some numbers (e.g., 0.875/2.050) appear in multiple rows. Please split the ablation into separate subtables with explicit descriptions of the baseline configuration.
  3. [Throughout] No error bars or repeated runs are reported for any experiment. Given the large claimed improvements, reporting mean and standard deviation over at least three training seeds for the main tables would strengthen the conclusions.
  4. [Section 3.3] There are typos, including "Denosing-Vision Transformer" (should be "Denoising-Vision Transformer") and "sythetic" in Section 3.2.
  5. [Table 2] For SeeingThroughFog, the paper should clarify how the ground-truth disparities were obtained and whether the same evaluation masks are used across all methods, since this dataset is known to have sparse and noisy LiDAR labels in fog.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the method trains on generated data and is evaluated on held-out real-world benchmarks, with no fitted parameter or self-citation chain forcing the reported gains.

full rationale

The paper's derivation chain is empirical rather than formal: it generates a synthetic adverse-weather stereo dataset (RST-Dataset) using Stable Diffusion/ControlNet conditioned on monocular depth, trains stereo networks with an L1 disparity loss against the original source disparity, and evaluates zero-shot on held-out DrivingStereo and SeeingThroughFog. No prediction is obtained by fitting to the test set; training and evaluation datasets are explicitly separated, and hyperparameters such as DDIM steps (50), patch matching n, and the encoder choice are fixed configurations rather than parameters tuned against the reported metrics. The ablation study varies the consistency module, data source, and encoder while keeping the evaluation fixed, which is the correct way to attribute gains. The paper's load-bearing assumption that generated image pairs preserve the original disparity field is a validity concern, not a circularity: the claim is empirical and could fail without making the derivation self-referential. Citations to prior work, including the authors' own [49], are used for components (depth estimation, diffusion, baselines) and are not invoked as uniqueness theorems or as the sole justification for the central claim. There is no equation in which an output is defined in terms of the target, no fitted parameter renamed as a prediction, and no self-citation chain that forces the result. Therefore the circularity score is 0.

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

No fitted mathematical parameter or derived constant is central to the method; the scores depend on empirical training. The free parameters listed are hand-set choices that could affect the generated data and the comparison. The axioms are domain assumptions about the fidelity of the generated stereo pairs and the fairness of the evaluation protocol.

free parameters (4)
  • top-n patch pairs in Disparity Fusion Module = not reported
    The consistency module keeps and fuses only the n most similar left-right patches; n is chosen by hand and no sensitivity study is given (Section 3.2, Equations 4-7).
  • weather prompt keywords = e.g., 'Rainy, dark clouds, wet pavement, raindrops, reflections, and misty air'
    Prompts are generated by an LLM and fixed per weather type; no prompt sensitivity analysis is provided (Section 3.2).
  • number of diffusion sampling steps = 50
    Fixed DDIM scheduler steps used for all generated images; standard but not ablated (Section 4.1).
  • number of training steps = not specified (equal across datasets)
    The paper states all models are trained for the same number of steps but does not report the number, so the fairness of dataset comparisons cannot be checked (Section 4.1).
assumptions (5)
  • domain assumption Stable Diffusion with DepthAnythingV2-conditioned ControlNet preserves the scene structure of the source image while applying adverse weather effects.
    Invoked in Section 3.2 to justify using the original disparity ground truth as supervision for generated images; never quantitatively verified.
  • domain assumption DepthAnythingV2 monocular depth estimates are accurate enough to guide generation for both left and right views.
    Used in Equation (1) as Dpred; depth errors would misalign the two generated views and corrupt stereo geometry.
  • domain assumption The patch-matching fusion module improves stereo consistency without distorting disparity values.
    Section 3.2, Equations (4)-(7); the authors provide qualitative images only, no epipolar error or disparity fidelity measurement.
  • domain assumption DrivingStereo and SeeingThroughFog are valid zero-shot testbeds for adverse-weather stereo matching.
    Section 4.1; the paper does not analyze label noise or sensor limitations in these real datasets.
  • domain assumption Comparisons across models trained on different datasets with equal training steps are meaningful indicators of data quality.
    Section 4.1; no learning curves or convergence checks are reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RobuSTereo: Robust Zero-Shot Stereo Matching under Adverse Weather." pith.science (2026). https://pith.science/paper/NSPGQLFW

@misc{pith2026250701653,
  author       = {Pith},
  title        = {Pith review of: RobuSTereo: Robust Zero-Shot Stereo Matching under Adverse Weather},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NSPGQLFW}},
  note         = {Machine review of arXiv:2507.01653}
}
read the original abstract

Learning-based stereo matching models struggle in adverse weather conditions due to the scarcity of corresponding training data and the challenges in extracting discriminative features from degraded images. These limitations significantly hinder zero-shot generalization to out-of-distribution weather conditions. In this paper, we propose \textbf{RobuSTereo}, a novel framework that enhances the zero-shot generalization of stereo matching models under adverse weather by addressing both data scarcity and feature extraction challenges. First, we introduce a diffusion-based simulation pipeline with a stereo consistency module, which generates high-quality stereo data tailored for adverse conditions. By training stereo matching models on our synthetic datasets, we reduce the domain gap between clean and degraded images, significantly improving the models' robustness to unseen weather conditions. The stereo consistency module ensures structural alignment across synthesized image pairs, preserving geometric integrity and enhancing depth estimation accuracy. Second, we design a robust feature encoder that combines a specialized ConvNet with a denoising transformer to extract stable and reliable features from degraded images. The ConvNet captures fine-grained local structures, while the denoising transformer refines global representations, effectively mitigating the impact of noise, low visibility, and weather-induced distortions. This enables more accurate disparity estimation even under challenging visual conditions. Extensive experiments demonstrate that \textbf{RobuSTereo} significantly improves the robustness and generalization of stereo matching models across diverse adverse weather scenarios.

Figures

Figures reproduced from arXiv: 2507.01653 by the authors.

Figure 1
Figure 1. Comparison of different stereo datasets. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. RobuSTereo Framework for Stereo Matching under Adverse Weather Conditions. The framework consists of three components: [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Visualization of generated data under adverse weather [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Qualitative results under adverse conditions. Our method produces more accurate and consistent disparity maps, preserving fine [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Comparison of Feature Maps from Different Encoders. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Qualitative visualization of point cloud results under adverse conditions. Point clouds generated by other methods show signif [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

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

  1. M3D-Stereo: A Multiple-Medium and Multiple-Degradation Dataset for Stereo Image Restoration

    cs.CV 2026-04 accept novelty 7.0 of 10

    M3D-Stereo supplies 7904 aligned stereo pairs across four multi-degradation scenarios with six progressive levels and pixel-consistent ground truths to benchmark image restoration and stereo matching.

  2. StereoPolicy: Improving Robotic Manipulation Policies via Stereo Perception

    cs.RO 2026-05 unverdicted novelty 6.0 of 10

    StereoPolicy fuses stereo image pairs via a Stereo Transformer on pretrained 2D encoders to boost robotic manipulation policies, showing gains over monocular, RGB-D, point cloud, and multi-view methods in simulations ...

  3. StereoPolicy: Improving Robotic Manipulation Policies via Stereo Perception

    cs.RO 2026-05 unverdicted novelty 4.0 of 10

    StereoPolicy fuses left-right image features via cross-attention to deliver consistent gains over RGB, RGB-D, point cloud, and multi-view baselines in simulation and real-robot manipulation tasks.

Reference graph

Works this paper leans on

52 extracted references · 50 canonical work pages · cited by 2 Pith papers

  1. [1]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ah- mad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint, 2023. 4

    work page 2023

  2. [2]

    Blended diffusion for text-driven editing of natural images

    Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Pro- ceedings of IEEE International Conference on Computer Vi- sion and Pattern Recognition, pages 18208–18218, 2022. 2

    work page 2022

  3. [3]

    Spatext: Spatio-textual representation for con- trollable image generation

    Omri Avrahami, Thomas Hayes, Oran Gafni, Sonal Gupta, Yaniv Taigman, Devi Parikh, Dani Lischinski, Ohad Fried, and Xi Yin. Spatext: Spatio-textual representation for con- trollable image generation. In Proceedings of IEEE Interna- tional Conference on Computer Vision and Pattern Recogni- tion, pages 18370–18380, 2023. 2

    work page 2023

  4. [4]

    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. In Proceedings of the IEEE Int. Conference on Intelligent Robots and Systems, 2021. 5, 6

    work page 2021

  5. [5]

    Multidiffusion: Fusing diffusion paths for controlled image generation

    Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. 2023. 3

    2023

  6. [6]

    Masksketch: Unpaired structure-guided masked image generation

    Dina Bashkirova, Jose Lezama, Kihyuk Sohn, Kate Saenko, and Irfan Essa. Masksketch: Unpaired structure-guided masked image generation. In Proceedings of IEEE Interna- tional Conference on Computer Vision and Pattern Recogni- tion, pages 1879–1889, 2023. 3

    work page 2023

  7. [7]

    Seeing through fog without seeing fog: Deep multimodal sensor fusion in unseen adverse weather

    Mario Bijelic, Tobias Gruber, Fahim Mannan, Florian Kraus, Werner Ritter, Klaus Dietmayer, and Felix Heide. Seeing through fog without seeing fog: Deep multimodal sensor fusion in unseen adverse weather. In Proceedings of IEEE International Conference on Computer Vision and Pattern Recognition, pages 11682–11692, 2020. 1, 2, 5, 6, 7

    work page 2020

  8. [8]

    In- structpix2pix: Learning to follow image editing instructions

    Tim Brooks, Aleksander Holynski, and Alexei A Efros. In- structpix2pix: Learning to follow image editing instructions. In Proceedings of IEEE International Conference on Com- puter Vision and Pattern Recognition , pages 18392–18402,

    work page

Show all 52 references
  1. [9]

    Vir- tual kitti 2, 2020

    Yohann Cabon, Naila Murray, and Martin Humenberger. Vir- tual kitti 2, 2020. 1, 2, 5, 6, 7, 8

    2020

  2. [10]

    Pyramid stereo matching network

    Jia-Ren Chang and Yong-Sheng Chen. Pyramid stereo matching network. In Proceedings of IEEE International Conference on Computer Vision and Pattern Recognition , pages 5410–5418, 2018. 1, 2, 5, 6, 7

    2018

  3. [11]

    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, 2025. 1, 5, 6

    2025

  4. [12]

    Diml/cvl rgb-d dataset: 2m rgb-d images of natural indoor and outdoor scenes

    Jaehoon Cho, Dongbo Min, Youngjung Kim, and Kwanghoon Sohn. Diml/cvl rgb-d dataset: 2m rgb-d images of natural indoor and outdoor scenes. arXiv preprint,

  5. [13]

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei- Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of IEEE International Conference on Computer Vision and Pattern Recognition, 2009. 5

    2009

  6. [14]

    Sid: Stereo image dataset for au- tonomous driving in adverse conditions

    Zaid A El-Shair, Abdalmalek Abu-raddaha, Aaron Cofield, Hisham Alawneh, Mohamed Aladem, Yazan Hamzeh, and Samir A Rawashdeh. Sid: Stereo image dataset for au- tonomous driving in adverse conditions. In NAECON 2024- IEEE National Aerospace and Electronics Conference, pages 403–...

    2024

  7. [15]

    Scene-based text-to-image generation with human priors,

    Oran Gafni, Adam Polyak, and Yaniv Nechemia Taigman. Scene-based text-to-image generation with human priors,

  8. [16]

    Virtual worlds as proxy for multi-object tracking anal- ysis

    Adrien Gaidon, Qiao Wang, Yohann Cabon, and Eleonora Vig. Virtual worlds as proxy for multi-object tracking anal- ysis. In Proceedings of IEEE International Conference on Computer Vision and Pattern Recognition , pages 4340– 4349, 2016. 1

    2016

  9. [17]

    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 Proceedings of IEEE International Conference on Computer Vision and Pattern Recognition , pages 3354– 3361, 2012. 2, 5

    2012

  10. [18]

    Cascade cost volume for high-resolution multi-view stereo and stereo matching

    Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. In Proceedings of IEEE International Conference on Computer Vision and Pat- tern Recognition, 2020. 5, 6

    2020

  11. [19]

    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 IEEE International Conference on Computer Vision and Pattern Recognition, pages 5459–5469, 2024. 5, 6

    2024

  12. [20]

    Camera-lidar cross-modality gait recognition

    Wenxuan Guo, Yingping Liang, Zhiyu Pan, Ziheng Xi, Jian- jiang Feng, and Jie Zhou. Camera-lidar cross-modality gait recognition. arXiv preprint, 2024. 1

    2024

  13. [21]

    Group-wise correlation stereo network

    Xiaoyang Guo, Kai Yang, Wukui Yang, Xiaogang Wang, and Hongsheng Li. Group-wise correlation stereo network. In Proceedings of IEEE International Conference on Computer Vision and Pattern Recognition, pages 3273–3282, 2019. 2, 5, 6

    2019

  14. [22]

    Openstereo: A comprehensive benchmark for stereo matching and strong baseline

    Xianda Guo, Juntao Lu, Chenming Zhang, Yiqi Wang, Yiqun Duan, Tian Yang, Zheng Zhu, and Long Chen. Openstereo: A comprehensive benchmark for stereo matching and strong baseline. arXiv preprint, 2023. 1, 2, 5, 6, 7, 8

    2023

  15. [23]

    Stereo anything: Unifying stereo matching with large- scale mixed data

    Xianda Guo, Chenming Zhang, Youmin Zhang, Dujun Nie, Ruilin Wang, Wenzhao Zheng, Matteo Poggi, and Long Chen. Stereo anything: Unifying stereo matching with large- scale mixed data. arXiv preprint, 2024. 1, 2, 5, 6, 7, 8

    2024

  16. [24]

    Light- stereo: Channel boost is all your need for efficient 2d cost aggregation

    Xianda Guo, Chenming Zhang, Youmin Zhang, Wenzhao Zheng, Dujun Nie, Matteo Poggi, and Long Chen. Light- stereo: Channel boost is all your need for efficient 2d cost aggregation. arXiv preprint, 2024. 5, 6, 7

    2024

  17. [25]

    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 IEEE International Confer- ence on Computer Vision and Pattern Recognition , pa...

    2024

  18. [26]

    End-to-end learning of geometry and context for deep stereo regression

    Alex Kendall, Hayk Martirosyan, Saumitro Dasgupta, Peter Henry, Ryan Kennedy, Abraham Bachrach, and Adam Bry. End-to-end learning of geometry and context for deep stereo regression. In Proceedings of IEEE International Conference on Computer Vision and Pattern Recognition, pag...

  19. [27]

    Vidtome: Video token merging for zero-shot video editing

    Xirui Li, Chao Ma, Xiaokang Yang, and Ming-Hsuan Yang. Vidtome: Video token merging for zero-shot video editing. In Proceedings of IEEE International Conference on Com- puter Vision and Pattern Recognition, 2024. 3

    2024

  20. [28]

    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. In Proceedings of IEEE International Conference on Computer V...

    2016

  21. [29]

    Object scene flow for autonomous vehicles

    Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In Proceedings of IEEE International Conference on Computer Vision and Pattern Recognition , pages 3061–3070, 2015. 2, 5, 6, 7, 8

    2015

  22. [30]

    Dinov2: Learning robust visual features without supervision

    Maxime Oquab, Timoth ´ee Darcet, Th ´eo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint, 2023. 7, 8

    2023

  23. [31]

    Slam2: an open-source slam system for monocular, stereo, and rgb-d cameras

    Mur-Artal R Tard ´os JD Orb. Slam2: an open-source slam system for monocular, stereo, and rgb-d cameras. IEEE Transactions on Robotics, 33(5):1255–1262, 2017. 1

    2017

  24. [32]

    High-resolution image syn- thesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models. In Proceedings of IEEE International Conference on Computer Vision and Pattern Recognition, pages 10684–10695, 2022. 2, 5

    2022

  25. [33]

    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. In Proceedings of IEEE International Conference on Computer Vision and Pattern Recognition, pages 4510–4520, 2018. 7, 8

    2018

  26. [34]

    The surprising effectiveness of diffusion models for optical flow and monocular depth estimation

    Saurabh Saxena, Charles Herrmann, Junhwa Hur, Abhishek Kar, Mohammad Norouzi, Deqing Sun, and David J Fleet. The surprising effectiveness of diffusion models for optical flow and monocular depth estimation. Proceedings of Ad- vances in Neural Information Processing Systems, 36...

    2024

  27. [35]

    Cfnet: Cascade and fused cost volume for robust stereo matching

    Zhelun Shen, Yuchao Dai, and Zhibo Rao. Cfnet: Cascade and fused cost volume for robust stereo matching. In Pro- ceedings of IEEE International Conference on Computer Vi- sion and Pattern Recognition, pages 13906–13915, 2021. 2, 5, 6

    2021

  28. [36]

    A tutorial on principal component analysis

    Jonathon Shlens. A tutorial on principal component analysis. arXiv preprint, 2014. 7

    2014

  29. [37]

    Simonyan and A

    K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. In Proceed- ings of International Conferenceon Learning Representa- tions, 2015. 5

    2015

  30. [38]

    Dif- fusion models for monocular depth estimation: Overcoming challenging conditions

    Fabio Tosi, Pierluigi Zama Ramirez, and Matteo Poggi. Dif- fusion models for monocular depth estimation: Overcoming challenging conditions. arXiv preprint, 2024. 3

    2024

  31. [39]

    Fadnet: A fast and accurate network for disparity estimation

    Qiang Wang, Shaohuai Shi, Shizhen Zheng, Kaiyong Zhao, and Xiaowen Chu. Fadnet: A fast and accurate network for disparity estimation. In Proceedings of the IEEE Int. Con- ference on Robotics and Automation, pages 101–107. IEEE,

  32. [40]

    Irs: A large naturalis- tic indoor robotics stereo dataset to train deep models for disparity and surface normal estimation

    Qiang Wang, Shizhen Zheng, Qingsong Yan, Fei Deng, Kaiyong Zhao, and Xiaowen Chu. Irs: A large naturalis- tic indoor robotics stereo dataset to train deep models for disparity and surface normal estimation. pages 1–6, 2021. 2

    2021

  33. [41]

    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 IEEE International Conference on Computer Vision and Pattern Recognition , pages 19701–19710, 2024. 5, 6

    2024

  34. [42]

    Structure-guided ranking loss for single image depth prediction

    Ke Xian, Jianming Zhang, Oliver Wang, Long Mai, Zhe Lin, and Zhiguo Cao. Structure-guided ranking loss for single image depth prediction. In Proceedings of IEEE Interna- tional Conference on Computer Vision and Pattern Recogni- tion, pages 611–620, 2020. 2

    2020

  35. [43]

    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 IEEE International Conference on Computer Vision and Pattern Recognition, pages 21919–21928, 2023. 1, 2, 5, 6, 7

    2023

  36. [44]

    Security and privacy of smart home systems based on the internet of things and stereo match- ing algorithms

    Aimin Yang, Chunying Zhang, Yongjie Chen, Yunxi Zhuan- sun, and Huixiang Liu. Security and privacy of smart home systems based on the internet of things and stereo match- ing algorithms. IEEE Internet of Things Journal, 7(4):2521– 2530, 2019. 1

    2019

  37. [45]

    Drivingstereo: A large-scale dataset for stereo matching in autonomous driving scenarios

    Guorun Yang, Xiao Song, Chaoqin Huang, Zhidong Deng, Jianping Shi, and Bolei Zhou. Drivingstereo: A large-scale dataset for stereo matching in autonomous driving scenarios. In Proceedings of IEEE International Conference on Com- puter Vision and Pattern Recognition, pages 899–...

    2019

  38. [46]

    Guibas, Dilip Krishnan, Kilian Q Weinberger, Yonglong Tian, and Yue Wang

    Jiawei Yang, Katie Z Luo, Jiefeng Li, Congyue Deng, Leonidas J. Guibas, Dilip Krishnan, Kilian Q Weinberger, Yonglong Tian, and Yue Wang. Dvt: Denoising vision trans- formers. 2024. 5

    2024

  39. [47]

    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, 2024. 4, 5

    2024

  40. [48]

    Ga-net: Guided aggregation net for end- to-end stereo matching

    Feihu Zhang, Victor Prisacariu, Ruigang Yang, and Philip HS Torr. Ga-net: Guided aggregation net for end- to-end stereo matching. In Proceedings of IEEE Interna- tional Conference on Computer Vision and Pattern Recogni- tion, pages 185–194, 2019. 2

    2019

  41. [49]

    Atlantis: En- abling underwater depth estimation with stable diffusion

    Fan Zhang, Shaodi You, Yu Li, and Ying Fu. Atlantis: En- abling underwater depth estimation with stable diffusion. In Proceedings of IEEE International Conference on Computer Vision and Pattern Recognition, pages 11852–11861, 2024. 3

    2024

  42. [50]

    Building a 3-d line-based map using stereo slam

    Guoxuan Zhang, Jin Han Lee, Jongwoo Lim, and Il Hong Suh. Building a 3-d line-based map using stereo slam. IEEE Transactions on Robotics, 31(6):1364–1377, 2015. 1

    2015

  43. [51]

    Adding conditional control to text-to-image diffusion models, 2023

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 3, 4, 5

    2023

  44. [2024]

    18/149,542

    US Patent App. 18/149,542. 2

Pith tools

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