Pith. sign in

REVIEW 4 major objections 5 minor 52 references

MonoMVSNet: Monocular Priors Guided Multi-View Stereo Network

T0 review · 4 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 Multi-view stereo struggles on flat, shiny, or edge pixels; MonoMVSNet fixes this by injecting monocular priors into the matching pipeline and achieves state-of-the-art DTU and Tanks-and-Temples results.

desk verdict Solid MVS paper with one load-bearing unit error in the depth-alignment equations that undermines the central dynamic-sampling claim; fixable, but not acceptable as written. read the letter →

arxiv 2507.11333 v1 pith:ZAU7MNBJ submitted 2025-07-15 cs.CV

Jianfei Jiang , Qiankun Liu , Haochen Yu , Hongyuan Liu , Liyong Wang , Jiansheng Chen , Huimin Ma This is my paper · ORCID
classification cs.CV
keywords multi-viewstereomonoculardepthpriorestimationcross-viewpositionencodingattentionmechanismsampling3Dreconstructionfoundationmodel
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

Multi-view stereo reconstructs depth by matching image features between viewpoints, but the matching itself breaks down on textureless regions, reflective surfaces, and object edges. This paper proposes MonoMVSNet, which injects priors from a monocular depth model—a network that estimates relative depth from a single image without any cross-view matching—into both the feature extraction and depth sampling stages of a cascade MVS network. It claims that fusing the reference view's monocular features into source views through a cross-view position encoding, and using aligned monocular depth to replace depth candidates at edge pixels, yields sharper depth at the exact regions where matching fails. The paper reports that this design achieves state-of-the-art overall error on the DTU dataset and the highest F-scores on the Tanks-and-Temples Intermediate and Advanced benchmarks, while using less GPU memory than comparable methods. If the claim holds, monocular priors become a cheap, effective remedy for MVS failure cases rather than a separate post-processing step.

What carries the argument

The argument is carried by three components. Cross-View Position Encoding (CVPE) warps reference and source features onto each other's depth-hypothesis planes, encodes camera intrinsics and extrinsics through an MLP and squeeze-and-excitation layer, and adds the resulting positional signal before intra-view and inter-view attention, so the attention mechanism sees 3D spatial relationships rather than flat 2D coordinates. Monocular depth alignment fits a linear model $a \hat{D}^{\mathrm{mono}}_s + b$ to the inverse of the previous-scale predicted depth on the most confident pixels (Eqs. 4–5), converting relative monocular depth into metric units; the aligned value is then swapped into the depth-candidate set at edge pixels. A relative consistency loss $L^{\mathrm{rc}}_3$ samples pixel pairs and penalizes any pair whose predicted depth differences have opposite sign to the aligned monocular depth differences, enforcing order agreement without requiring absolute metric supervision from the monocular model.

What would settle it

Take a scene with a wide depth range where the confident pixels used for the fit lie mostly on the background, then check whether the aligned monocular depth inserted at edge pixels anywhere in the image falls inside the true local depth interval. A single global scale-and-shift fit cannot represent spatially varying scale, so such a scene should expose misaligned candidates; if edge-region depth error on this scene is not better than the uniform inverse-depth baseline, the dynamic-sampling claim is refuted.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that a pre-trained monocular depth model can be treated as a plug-in prior for multi-view stereo. MonoMVSNet uses the monocular feature of the reference view only, adding it to the FPN feature at the coarsest scale and then propagating it to source-view features through attention blocks whose position encoding is computed by warping features across views and embedding camera parameters—the proposed Cross-View Position Encoding. For depth, the monocular map is aligned to the network's own previous-scale depth by a least-squares fit in inverse-depth space (Eqs. 4–5) and the aligned value is substituted for the nearest inverse-depth candidate at pixels flagged by an edge detector, so that depth hypotheses near discontinuities come from monocular reasoning rather than uniform sampling. A relative consistency loss (Eqs. 8–9) then supervises the final-scale probability volume by penalizing pairs of pixels whose predicted depth ranks contradict the aligned monocular depth ranks. The combination is reported to produce the best overall point-cloud error on DTU (0.278 with five views) and first place on both Tanks-and-Temples subsets, with 2.01 GB inference memory at 832×1152.

Load-bearing premise

The load-bearing premise is that the aligned monocular depth, obtained by fitting a single linear model to the inverse of the network's previous-scale depth, is accurate enough in metric units to be used as a replacement depth candidate at edge pixels; if this fit is off, the dynamic sampling injects systematically wrong candidates at exactly the discontinuities it is meant to fix.

Editorial extensions

If this is right

  • If correct, MVS networks can inherit the robustness of monocular depth in textureless and reflective regions without paying the cost of running a monocular model on every input view, since only the reference view is processed.
  • The dynamic depth sampling replaces uniform inverse-depth candidates at edges with aligned monocular values, which should specifically improve completeness at depth discontinuities; the reported DTU completeness gain from 0.282 to 0.248 supports that.
  • The relative consistency loss offers a way to supervise depth ranking using relative priors, which could be applied in settings where absolute ground-truth depth is unavailable or noisy.
  • The comparison with a variant that extracts monocular features from all views shows that reference-only extraction not only cuts memory by about 65% but also improves accuracy, implying that monocular features act as a corrective signal rather than a feature source that needs to be replicated everywhere.
  • On Tanks-and-Temples, the method generalizes from DTU training to large-scale outdoor scenes, suggesting the monocular priors help bridge the domain gap in addition to fixing local matching failures.

Reading between the lines

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

  • Beyond the paper's claims: because the alignment is a single global scale-and-shift fit, the method implicitly assumes the monocular relative depth is metrically consistent across the whole image; a piecewise or per-pixel alignment could extend the benefit to scenes with wide depth variation, but that is not tested here.
  • The success of replacing depth candidates only at detected edge pixels suggests a testable interaction: if the edge detector were replaced by an uncertainty or reliability map from the monocular model, the dynamic sampling could be applied where the model is most confident, which the paper does not explore.
  • The same recipe—reference-view monocular feature fusion plus aligned-depth candidate injection—could be transferred to other cascade MVS backbones or to video-based 3D reconstruction, since the modules are decoupled from the cost-volume regularizer; the paper only demonstrates it inside MonoMVSNet.
  • A direct consequence of the reference-only design is that improvements in the underlying monocular foundation model should translate into accuracy gains for MVS at zero extra training cost; the supplementary experiments replacing the monocular backbone with several alternative pre-trained models support this direction but only test one backbone.
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

4 major / 5 minor

Summary. MonoMVSNet integrates features and depth from a pretrained monocular foundation model (Depth Anything V2) into a multi-view stereo network. The reference-view monocular feature is fused with FPN features and propagated to source views via an attention mechanism with a newly designed cross-view position encoding. The monocular depth is aligned to the network's coarse depth and then used to replace depth candidates for edge pixels during inverse depth sampling, and a relative consistency loss supervises the predicted depth against the monocular depth. Experiments report state-of-the-art point cloud metrics on DTU and Tanks-and-Temples, with lower GPU memory than several recent MVS methods.

Significance. The paper proposes a plausible integration of monocular depth and features into a cascade MVS pipeline. If the empirical results are reproducible, the method is a useful engineering contribution: it achieves strong scores on DTU and Tanks-and-Temples while using less GPU memory than several recent MVS methods. The authors provide code, an extensive ablation study, and a clear efficiency analysis, which are strengths. However, the dynamic depth sampling mechanism is not coherently described due to the unit mismatch in Eqs. (4)-(5), and the SOTA claims need more rigorous statistical support.

major comments (4)
  1. [Section 3.3, Eqs. (4)-(5)] The alignment in Eq. (4) fits a linear model (a * Dhat_mono_s + b) to the inverse of the previous-scale predicted depth 1/Dhat_s. Equation (5) then defines Dalign_s = a * Dhat_mono_s + b, which is an estimator of inverse metric depth (units 1/mm on DTU, where the depth range is 425-935 mm). The subsequent paragraph states that Dalign_s[c] is compared with and replaces the metric depth candidates d_{i,s} during inverse depth sampling. This is a unit mismatch: on DTU the aligned value would be roughly 0.001-0.002, whereas the candidates are hundreds of mm, so injecting Dalign_s as a depth candidate would place it far outside the valid range and break the winner-take-all selection. If the candidates are instead intended to be inverse depth, then Eq. (6) would produce inverse depth and the paper does not describe the required inversion to metric depth. Either interpretation makes the dynamic depth sampling mechanism internally inconsistent as written; please clarify the exact quantity and the required transformations.
  2. [Section 3.3, alignment filtering] The alignment parameters (a,b) are estimated from the top 80% of pixels by confidence from the previous scale, but the dynamic sampling applies the aligned depth to edge pixels, which are exactly the pixels likely to have low confidence and thus to be excluded from the fit. Using a global scale/shift estimated on non-edge regions for edge pixels is not justified, and the claimed benefit of 'fine-grained relative depth information' in edge regions is not supported by the described procedure. Please provide evidence (e.g., an ablation that fits the alignment on edge pixels as well) or revise the procedure.
  3. [Section 4.3, Tables 1-2] The state-of-the-art claim rests on improvements that are small relative to the expected variability (e.g., DTU overall 0.278 vs. 0.281 for MVSFormer++; Tanks-and-Temples intermediate 68.63 vs. 67.18). No error bars, multiple seeds, or statistical significance tests are reported. Given that the point cloud metrics depend on fusion parameters, please report variance across at least three runs or a sensitivity analysis to support the claim of consistent improvement.
  4. [Abstract and Section 4.3] The abstract and Section 4.3 state that MonoMVSNet 'ranks first on the Tanks-and-Temples Intermediate and Advanced benchmarks,' but the comparison in Table 2 is against a selected subset of learning-based methods. The official Tanks-and-Temples leaderboard contains many other published entries, and the paper does not show that the reported scores surpass all of them. Please either qualify the claim to 'among the compared methods' or provide a full comparison against the official leaderboard.
minor comments (5)
  1. [Section 3.5] The phrase 'the the depth produced' contains a duplicated determiner and should read 'the depth produced'.
  2. [Section 3.3] The sentence 'we adopt an inverse depth sampling strategy to sampled depth candidates' should read 'to sample depth candidates'.
  3. [Table 3 caption vs. Section 4.4] Table 3 states 'using normal fusion strategy' while Section 4.4 states 'the dynamic fusion strategy [38]' is used; please reconcile this inconsistency.
  4. [Section 2, Related Work] The text mentions 'MaG-Net [22] samples based on the monocular depth probability distribution,' but reference [22] is GBi-Net and no MaG-Net reference is listed; please add the missing citation or correct the text.
  5. [Equation (9)] Equation (9) is typeset with the multiplication dot after the line break and unbalanced parentheses; the formula should be written as em = (Dprob_s[c1_m] - Dprob_s[c2_m]) * Sign(Dhat_mono_s[c1_m] - Dhat_mono_s[c2_m]).

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: monocular priors are external, and the aligned monocular depth is only an auxiliary candidate, not the network's prediction target.

full rationale

The paper's derivation chain is self-contained with respect to the benchmark claims. The monocular feature and depth priors come from an externally pretrained model (Depth Anything V2, [41]), not from the DTU or Tanks-and-Temples evaluation data. The alignment in Eq. (4)-(5) fits scale and shift parameters to the network's own coarse depth prediction from the previous scale, and the result is used only to replace a single depth candidate at selected edge pixels during sampling; the final depth in Eq. (6) is still selected from the cost-volume probability volume and is supervised by a cross-entropy loss against ground truth. The relative consistency loss (Eqs. 8-9) encourages order agreement with monocular depth but does not force the predicted depth to equal the monocular depth, and it is applied only at the final scale with a modest weight. No parameter is fitted to the benchmark ground truth and then reported as a prediction, and no load-bearing claim rests on a self-citation. The unit-mismatch concern about Dalign being used as a metric depth candidate is a correctness risk, not circularity, because the aligned value is not, by construction, the network's output. Self-citations to the authors' prior MVS works appear only in the introduction/related work and are not used to justify the central mechanism. The reported gains are therefore externally testable rather than forced by the paper's definitions.

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

The method relies on a pre-trained monocular foundation model and an edge detection network from prior work, without re-deriving their reliability. Several hand-chosen hyperparameters govern the depth sampling and loss weighting. No new physical entities are introduced.

free parameters (5)
  • edge threshold lambda = not specified
    Pixels with edge confidence above lambda are selected for monocular-guided depth candidate replacement (Section 3.3).
  • top confidence fraction for alignment = 80%
    The top 80% highest-confidence pixels are used to fit the monocular depth alignment in Section 3.3.
  • relative consistency loss weight gamma = not specified
    Weight on the relative consistency loss in Eq. 10.
  • depth candidate counts and intervals = 8-8-4-4, 0.5
    Coarse-to-fine depth hypothesis counts and interval hyperparameters used in training (Section 4.2).
  • number of random samples M = not specified
    Number of pixel pairs sampled for the relative consistency loss in Eq. 8.
assumptions (4)
  • domain assumption Pre-trained monocular model (Depth Anything V2) provides reliable relative depth and useful features in challenging regions.
    The method's three contributions all rely on this assumed quality; the paper does not analyze failure cases of the monocular prior.
  • domain assumption Lightweight edge detection network [29] produces a reliable edge confidence map.
    The dynamic depth sampling uses these edge maps to decide where to inject monocular depth.
  • domain assumption The confidence map from the previous scale identifies reliable pixels for alignment.
    Section 3.3 keeps the top 80% confidence pixels for the least-squares fit without validating this filtering rule.
  • standard math Standard multi-view projective geometry and inverse depth sampling are applicable.
    Eq. 2 uses the usual pinhole camera projection, and the inverse depth strategy is adopted from MVSTER [32].

how reviews work

0 comments
Cite this review

Pith. "Pith review of MonoMVSNet: Monocular Priors Guided Multi-View Stereo Network." pith.science (2026). https://pith.science/paper/ZAU7MNBJ

@misc{pith2026250711333,
  author       = {Pith},
  title        = {Pith review of: MonoMVSNet: Monocular Priors Guided Multi-View Stereo Network},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZAU7MNBJ}},
  note         = {Machine review of arXiv:2507.11333}
}
read the original abstract

Learning-based Multi-View Stereo (MVS) methods aim to predict depth maps for a sequence of calibrated images to recover dense point clouds. However, existing MVS methods often struggle with challenging regions, such as textureless regions and reflective surfaces, where feature matching fails. In contrast, monocular depth estimation inherently does not require feature matching, allowing it to achieve robust relative depth estimation in these regions. To bridge this gap, we propose MonoMVSNet, a novel monocular feature and depth guided MVS network that integrates powerful priors from a monocular foundation model into multi-view geometry. Firstly, the monocular feature of the reference view is integrated into source view features by the attention mechanism with a newly designed cross-view position encoding. Then, the monocular depth of the reference view is aligned to dynamically update the depth candidates for edge regions during the sampling procedure. Finally, a relative consistency loss is further designed based on the monocular depth to supervise the depth prediction. Extensive experiments demonstrate that MonoMVSNet achieves state-of-the-art performance on the DTU and Tanks-and-Temples datasets, ranking first on the Tanks-and-Temples Intermediate and Advanced benchmarks. The source code is available at https://github.com/JianfeiJ/MonoMVSNet.

Figures

Figures reproduced from arXiv: 2507.11333 by the authors.

Figure 1
Figure 1. Row 1: (a) Comparison with SOTA methods in terms of overall error and GPU memory consumption at a resolution of 832×1152 with 5-view images on the DTU [1] test set, where lower is better; (b) Comparison with SOTA methods on the Tanks-and-Temples [17] benchmark, where higher is better. Row 2-3: Qualitative depth comparison with Depth Anything V2 [41], MVSFormer [3], and MVSFormer++ [4] on scan13 from DTU test set. Ou… view at source ↗
Figure 2
Figure 2. Overview of the proposed MonoMVSNet. (1) Exploitation of monocular feature: The reference monocular feature, extracted by the mono encoder model, is used to enhance the reference FPN feature and integrated into source features by attention mechanism with cross-view position encoding. (2) Exploitation of monocular depth: The monocular depth of the reference image, output by the mono decoder, is aligned to guide the d… view at source ↗
Figure 3
Figure 3. Illustration of the Cross-View Position Encoding (CVPE). Each pair of reference and source features is warped into the respective views. Together with the camera parameters, they undergo camera embedding to generate the CVPE for both the ref￾erence and source features. reference feature to source features more effectively. Supposing the depth hypotheses in the 0-th scale are {di,0} D0−1 i=0 , the n-th source feature… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison of predicted depth maps with GeoMVSNet [ [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Qualitative comparison of reconstructed point clounds with ET-MVSNet [ [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: Additional depth maps visualization comparing the Baseline, MonoMVSNet, Ground Truth, Depth Anything V2 (DAV2), and [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: More visualization results of all reconstructed point clouds on the DTU [ [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: More visualization results of all reconstructed point clouds on the Tanks-and-Temples [ [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

52 extracted references · 44 canonical work pages

  1. [1]

    Large-scale data for multiple-view stereopsis

    Henrik Aanæs, Rasmus Ramsbøl Jensen, George V ogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. International Journal of Computer Vision, 120:153–168, 2016. 1, 6, 7, 12, 14, 15

    work page 2016

  2. [2]

    Depth pro: Sharp monocular metric depth in less than a second

    Aleksei Bochkovskii, Ama ¨el Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. arXiv preprint arXiv:2410.02073, 2024. 12

    arXiv 2024

  3. [3]

    Mvsformer: Multi-view stereo by learning robust image features and temperature-based depth

    Chenjie Cao, Xinlin Ren, and Yanwei Fu. Mvsformer: Multi-view stereo by learning robust image features and temperature-based depth. Transactions on Machine Learn- ing Research, 2022. 1, 2, 6, 7, 8

    work page 2022

  4. [4]

    Mvsformer++: Revealing the devil in transformer’s details for multi-view stereo

    Chenjie Cao, Xinlin Ren, and Yanwei Fu. Mvsformer++: Revealing the devil in transformer’s details for multi-view stereo. arXiv preprint arXiv:2401.11673, 2024. 1, 2, 6, 7, 8, 13

    arXiv 2024

  5. [5]

    Deep stereo using adap- tive thin volume representation with uncertainty awareness

    Shuo Cheng, Zexiang Xu, Shilin Zhu, Zhuwen Li, Li Erran Li, Ravi Ramamoorthi, and Hao Su. Deep stereo using adap- tive thin volume representation with uncertainty awareness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2524–2534, 2020. 2

    work page 2020

  6. [6]

    Twins: Revisiting the design of spatial attention in vision transformers

    Xiangxiang Chu, Zhi Tian, Yuqing Wang, Bo Zhang, Haib- ing Ren, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Twins: Revisiting the design of spatial attention in vision transformers. Advances in neural information processing systems, 34:9355–9366, 2021. 2

    work page 2021

  7. [7]

    Deformable convolutional networks

    Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In Proceedings of the IEEE International Confer- ence on Computer Vision, pages 764–773, 2017. 2

    work page 2017

  8. [8]

    Transmvs- net: Global context-aware multi-view stereo network with transformers

    Yikang Ding, Wentao Yuan, Qingtian Zhu, Haotian Zhang, Xiangyue Liu, Yuanjiang Wang, and Xiao Liu. Transmvs- net: Global context-aware multi-view stereo network with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 8585– 8594, 2022. 1, 2, 4, 6, 7

    work page 2022

Show all 52 references
  1. [9]

    Massively parallel multiview stereopsis by surface normal diffusion

    Silvano Galliani, Katrin Lasinger, and Konrad Schindler. Massively parallel multiview stereopsis by surface normal diffusion. In Proceedings of the IEEE International Confer- ence on Computer Vision, pages 873–881, 2015. 1, 6, 7

    2015

  2. [10]

    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 the IEEE/CVF conference on computer vision and pattern recognition, pages 2495–2504, 2020. 1, 2, 6, 7

    2020

  3. [11]

    Group-wise correlation stereo network

    Xiaoyang Guo, Kai Yang, Wukui Yang, Xiaogang Wang, and Hongsheng Li. Group-wise correlation stereo network. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 3273–3282, 2019. 5

    2019

  4. [12]

    Squeeze-and-excitation net- works

    Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation net- works. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 7132–7141, 2018. 4, 12

    2018

  5. [13]

    Di-mvs: Learning efficient multi-view stereo with depth- aware iterations

    Jianfei Jiang, Mingwei Cao, Jun Yi, and Chenglong Li. Di-mvs: Learning efficient multi-view stereo with depth- aware iterations. In ICASSP 2024 - 2024 IEEE Interna- tional Conference on Acoustics, Speech and Signal Process- ing (ICASSP), pages 3180–3184, 2024. 1

    2024

  6. [14]

    Rrt-mvs: Recurrent regular- ization transformer for multi-view stereo

    Jianfei Jiang, Liyong Wang, Haochen Yu, Tianyu Hu, Jian- sheng Chen, and Huimin Ma. Rrt-mvs: Recurrent regular- ization transformer for multi-view stereo. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3994– 4002, 2025. 1

    2025

  7. [15]

    Parallel feature pyra- mid network for object detection

    Seung-Wook Kim, Hyong-Keun Kook, Jee-Young Sun, Mun-Cheon Kang, and Sung-Jea Ko. Parallel feature pyra- mid network for object detection. In Proceedings of the Eu- ropean Conference on Computer Vision (ECCV), pages 234– 250, 2018. 1, 2

    2018

  8. [16]

    Adam: A method for stochastic optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 ,

    arXiv

  9. [17]

    Tanks and temples: Benchmarking large-scale scene reconstruction

    Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG) , 36 (4):1–13, 2017. 1, 6, 7, 12, 13, 16

    2017

  10. [18]

    Learning deformable hypothesis sampling for ac- curate patchmatch multi-view stereo

    Hongjie Li, Yao Guo, Xianwei Zheng, and Hanjiang Xiong. Learning deformable hypothesis sampling for ac- curate patchmatch multi-view stereo. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3082– 3090, 2024. 6, 7

    2024

  11. [19]

    Wt-mvsnet: window-based transformers for multi-view stereo

    Jinli Liao, Yikang Ding, Yoli Shavit, Dihe Huang, Shihao Ren, Jia Guo, Wensen Feng, and Kai Zhang. Wt-mvsnet: window-based transformers for multi-view stereo. Advances in Neural Information Processing Systems , 35:8564–8576,

  12. [20]

    Protocar: Learning 3d vehicle prototypes from single-view and unconstrained driving scene images

    Hongyuan Liu, Haochen Yu, Bochao Zou, Juntao Lyu, Qi Mei, Jiansheng Chen, and Huimin Ma. Protocar: Learning 3d vehicle prototypes from single-view and unconstrained driving scene images. In Proceedings of the AAAI Confer- ence on Artificial Intelligence, pages 5460–5468, 2025. 1

    2025

  13. [21]

    When epipolar constraint meets non-local oper- ators in multi-view stereo

    Tianqi Liu, Xinyi Ye, Weiyue Zhao, Zhiyu Pan, Min Shi, and Zhiguo Cao. When epipolar constraint meets non-local oper- ators in multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18088– 18097, 2023. 2, 4, 6, 7, 8, 13

    2023

  14. [22]

    Generalized binary search network for highly-efficient multi-view stereo

    Zhenxing Mi, Chang Di, and Dan Xu. Generalized binary search network for highly-efficient multi-view stereo. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12991–13000, 2022. 2, 3

    2022

  15. [23]

    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 arXiv:2304.07193, 2023. 2, 12

    2023 arXiv

  16. [24]

    Pytorch: An im- perative style, high-performance deep learning library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming 9 Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An im- perative style, high-performance deep learning library. Ad- vances in neural information processing system...

    2019

  17. [25]

    Rethinking depth estimation for multi- view stereo: A unified representation

    Rui Peng, Rongjie Wang, Zhenyu Wang, Yawen Lai, and Ronggang Wang. Rethinking depth estimation for multi- view stereo: A unified representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8645–8654, 2022. 6, 7

    2022

  18. [26]

    Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer

    Ren ´e Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 44(3):1623–1637, 2020. 3

    2020

  19. [27]

    Vi- sion transformers for dense prediction

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

    2021

  20. [28]

    Pixelwise view selection for unstructured multi-view stereo

    Johannes L Sch ¨onberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Nether- lands, October 11-14, 2016, Proceedings, Part III 14, pag...

    2016

  21. [29]

    Tiny and efficient model for the edge detec- tion generalization

    Xavier Soria, Yachuan Li, Mohammad Rouhani, and An- gel D Sappa. Tiny and efficient model for the edge detec- tion generalization. In Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision, pages 1364–1373,

  22. [30]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko- reit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2

    2017

  23. [31]

    Is-mvsnet: importance sampling-based mvsnet

    Likang Wang, Yue Gong, Xinjun Ma, Qirui Wang, Kaixuan Zhou, and Lei Chen. Is-mvsnet: importance sampling-based mvsnet. In European Conference on Computer Vision, pages 668–683. Springer, 2022. 3

    2022

  24. [32]

    Mvster: Epipo- lar transformer for efficient multi-view stereo

    Xiaofeng Wang, Zheng Zhu, Guan Huang, Fangbo Qin, Yun Ye, Yijia He, Xu Chi, and Xingang Wang. Mvster: Epipo- lar transformer for efficient multi-view stereo. In European Conference on Computer Vision , pages 573–591. Springer,

  25. [33]

    Aa-rmvsnet: Adaptive aggregation recurrent multi-view stereo network

    Zizhuang Wei, Qingtian Zhu, Chen Min, Yisong Chen, and Guoping Wang. Aa-rmvsnet: Adaptive aggregation recurrent multi-view stereo network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6187– 6196, 2021. 2

    2021

  26. [34]

    Gomvs: Geometrically consistent cost aggregation for multi-view stereo

    Jiang Wu, Rui Li, Haofei Xu, Wenxun Zhao, Yu Zhu, Jinqiu Sun, and Yanning Zhang. Gomvs: Geometrically consistent cost aggregation for multi-view stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20207–20216, 2024. 6, 7, 8, 13

    2024

  27. [35]

    Planar prior assisted patch- match multi-view stereo

    Qingshan Xu and Wenbing Tao. Planar prior assisted patch- match multi-view stereo. In Proceedings of the AAAI Con- ference on Artificial Intelligence, pages 12516–12523, 2020. 1

    2020

  28. [36]

    Learning inverse depth re- gression for multi-view stereo with correlation cost volume

    Qingshan Xu and Wenbing Tao. Learning inverse depth re- gression for multi-view stereo with correlation cost volume. In Proceedings of the AAAI conference on artificial intelli- gence, pages 12508–12515, 2020. 5, 7

    2020

  29. [37]

    Multi-scale geometric consistency guided and planar prior assisted multi-view stereo

    Qingshan Xu, Weihang Kong, Wenbing Tao, and Marc Polle- feys. Multi-scale geometric consistency guided and planar prior assisted multi-view stereo. IEEE Transactions on Pat- tern Analysis and Machine Intelligence , 45(4):4945–4963,

  30. [38]

    Dense hybrid recurrent multi-view stereo net with dy- namic consistency checking

    Jianfeng Yan, Zizhuang Wei, Hongwei Yi, Mingyu Ding, Runze Zhang, Yisong Chen, Guoping Wang, and Yu-Wing Tai. Dense hybrid recurrent multi-view stereo net with dy- namic consistency checking. In European conference on computer vision, pages 674–689. Springer, 2020. 2, 7

    2020

  31. [39]

    Cost volume pyramid based depth inference for multi-view stereo

    Jiayu Yang, Wei Mao, Jose M Alvarez, and Miaomiao Liu. Cost volume pyramid based depth inference for multi-view stereo. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4877–4886,

  32. [40]

    Depth anything: Unleashing the power of large-scale unlabeled data

    Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371–10381, 2024. 2, 3, 12

    2024

  33. [41]

    Depth any- thing v2

    Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiao- gang Xu, Jiashi Feng, and Hengshuang Zhao. Depth any- thing v2. arXiv preprint arXiv:2406.09414, 2024. 1, 2, 3, 4, 12

    2024 arXiv

  34. [42]

    Mvsnet: Depth inference for unstructured multi-view stereo

    Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vi- sion (ECCV), pages 767–783, 2018. 1, 2, 6

    2018

  35. [43]

    Recurrent mvsnet for high-resolution multi-view stereo depth inference

    Yao Yao, Zixin Luo, Shiwei Li, Tianwei Shen, Tian Fang, and Long Quan. Recurrent mvsnet for high-resolution multi-view stereo depth inference. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5525–5534, 2019. 1, 2

    2019

  36. [44]

    Blendedmvs: A large- scale dataset for generalized multi-view stereo networks

    Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. Blendedmvs: A large- scale dataset for generalized multi-view stereo networks. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 1790–1799, 2020. 6

    2020

  37. [45]

    Constraining depth map geometry for multi-view stereo: A dual-depth approach with saddle- shaped depth cells

    Xinyi Ye, Weiyue Zhao, Tianqi Liu, Zihao Huang, Zhiguo Cao, and Xin Li. Constraining depth map geometry for multi-view stereo: A dual-depth approach with saddle- shaped depth cells. InProceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, pages 17661–17670,

  38. [46]

    Get3dgs: Generate 3d gaussians based on points deformation fields

    Haochen Yu, Weixi Gong, Jiansheng Chen, and Huimin Ma. Get3dgs: Generate 3d gaussians based on points deformation fields. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 1

    2024

  39. [47]

    Msp-mvs: Multi- granularity segmentation prior guided multi-view stereo

    Zhenlong Yuan, Cong Liu, Fei Shen, Zhaoxin Li, Jinguo Luo, Tianlu Mao, and Zhaoqi Wang. Msp-mvs: Multi- granularity segmentation prior guided multi-view stereo. In Proceedings of the AAAI Conference on Artificial Intelli- gence, pages 9753–9762, 2025. 1

    2025

  40. [48]

    Dvp-mvs: Synergize 10 depth-edge and visibility prior for multi-view stereo

    Zhenlong Yuan, Jinguo Luo, Fei Shen, Zhaoxin Li, Cong Liu, Tianlu Mao, and Zhaoqi Wang. Dvp-mvs: Synergize 10 depth-edge and visibility prior for multi-view stereo. In Pro- ceedings of the AAAI Conference on Artificial Intelligence , pages 9743–9752, 2025

    2025

  41. [49]

    Sed-mvs: Segmentation-driven and edge- aligned deformation multi-view stereo with depth restoration and occlusion constraint

    Zhenlong Yuan, Zhidong Yang, Yujun Cai, Kuangxin Wu, Mufan Liu, Dapeng Zhang, Hao Jiang, Zhaoxin Li, and Zhaoqi Wang. Sed-mvs: Segmentation-driven and edge- aligned deformation multi-view stereo with depth restoration and occlusion constraint. IEEE Transactions on Circuits and...

    2025

  42. [50]

    Multi-view stereo representation revist: Region-aware mvsnet

    Yisu Zhang, Jianke Zhu, and Lixiang Lin. Multi-view stereo representation revist: Region-aware mvsnet. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 17376–17385, 2023. 6, 7

    2023

  43. [51]

    Ge- omvsnet: Learning multi-view stereo with geometry percep- tion

    Zhe Zhang, Rui Peng, Yuxi Hu, and Ronggang Wang. Ge- omvsnet: Learning multi-view stereo with geometry percep- tion. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 21508–21518,

  44. [52]

    Sam2object: Consolidating view consistency via sam2 for zero-shot 3d instance segmentation

    Jihuai Zhao, Junbao Zhuo, Jiansheng Chen, and Huimin Ma. Sam2object: Consolidating view consistency via sam2 for zero-shot 3d instance segmentation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 19325–19334, 2025. 1 11 MonoMVSNet: Monocular Pr...

    2025

Pith tools

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