Pith. sign in

REVIEW 3 major objections 5 minor 46 references

Aerial Multi-View Stereo via Adaptive Depth Range Inference and Normal Cues

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Per-pixel depth ranges predicted from monocular depth and normal cues improve aerial multi-view stereo accuracy across three benchmarks.

desk verdict A solid, incremental aerial MVS paper: the depth-range predictor idea is genuinely new and well-ablated, but the SOTA claim rests on sub-centimeter margins without error bars or code. read the letter →

arxiv 2506.05655 v1 pith:QU7XNFAG submitted 2025-06-06 cs.CV eess.IV

Yimei Liu , Yakun Ju , Yuan Rao , Hao Fan , Junyu Dong , Feng Gao , Qian Du This is my paper
classification cs.CVeess.IV
keywords aerialmulti-viewstereoadaptivedepthrangepredictionmonocularpriornormalnormal-guidedcostaggregationrefinementremotesensing3Dreconstruction
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

ADR-MVS claims that aerial multi-view stereo fails when every pixel is searched over the same depth interval, because aerial views compress motion along epipolar lines and make feature matching ambiguous. Its answer is to let each pixel search its own depth range, inferred from two monocular cues: a relative depth estimate and a surface-normal estimate. The depth-range predictor widens the search where the cues disagree or geometry changes sharply, then narrows it in later cascade stages for precise regression. The same normal cue also guides cost aggregation and final depth refinement. On three aerial benchmarks the method reports the best mean absolute depth error while keeping runtime and memory near the lightest close-range baseline.

What carries the argument

The Depth Range Predictor (DRP) is the load-bearing component. It converts monocular depth into a normal map, extracts features from that map and from the monocular normal map, and runs a per-pixel cross-attention discrepancy-perception module in both feature orders: subtracting the shared information attended from one feature set leaves a discrepancy feature, which is regressed to a range map $R_{pred}$. Each cascade stage builds its depth candidate volume centered on the current depth map with radius $R_{pred}$, so the first stage can expand beyond predefined boundaries and later stages contract. Two further modules reuse the same monocular normal: Normal-Guided Cost Aggregation (NCA) weights geometrically relevant neighboring costs using normal-based depth ratios and feature-similarity weights, and Normal-Guided Depth Refinement (NDR) upsamples the half-resolution depth map with normal-guided residuals instead of RGB-guided ones.

What would settle it

On a held-out aerial dataset with ground-truth depth, compute the fraction of pixels whose true depth falls inside the first-stage predicted range and compare it with the same fraction for a fixed doubled range; if the learned range has no higher coverage, the accuracy gain is not coming from better range placement. A second check is to rotate the monocular normal field by a known angle and measure the change in mean absolute error, since a sharp degradation would show the pipeline leans on prior calibration rather than on learned geometric cues.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that the depth range used to build the cost volume, not just the architecture, is the main bottleneck for aerial matching. A per-pixel range map derived from the discrepancy between depth-to-normal and monocular normal features gives the first stage a wider, geometry-aware search space; later stages shrink the range around the regressed depth. The paper shows this by replacing the fixed SfM-derived range in a cascade with predicted range maps, plus normal-guided aggregation and refinement. The decisive comparison is on the WHU three-view test, where the method reports a mean absolute error of 9.4 cm against 12.4 cm for the close-range baseline and 9.7 cm for the previous best edge-guided aerial method, and it also achieves the lowest MAE on LuoJia-MVS and the cross-dataset München test.

Load-bearing premise

The whole approach depends on the frozen monocular depth and normal models being reliable enough that their disagreement marks where matching needs a wider search; if those priors are systematically miscalibrated on low-texture aerial scenes, the predicted ranges can miss the true depth and no later stage can fix it.

Editorial extensions

If this is right

  • The reported ablation indicates that the gain comes from replacing the fixed initial depth range with a learned per-pixel range: simply doubling the fixed range helps, tripling it hurts, while the predicted range beats both.
  • Monocular depth and normal priors contribute in complementary ways: removing either prior degrades accuracy, and the normal prior carries more weight in cost aggregation and refinement, where depth-computed normals are less stable.
  • The method transfers across monocular priors: substituting different depth and normal estimators still outperforms the baseline, suggesting the range-prediction mechanism rather than any single prior is what matters.
  • Efficiency follows from the design: replacing standard 3D convolution with normal-guided aggregation and adding lightweight modules keeps memory and runtime below the multi-branch edge-guided rival while improving accuracy.

Reading between the lines

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

  • If the discrepancy between depth-derived normals and monocular normals is the real signal, the same range-prediction idea could be applied to satellite stereo and high-altitude oblique imagery, where epipolar compression is even stronger.
  • The paper's visualizations show first-stage range maps correlating with monocular depth error; verifying that correlation on unseen sites would separate the DRP's benefit from the mere effect of a wider search range.
  • The reported finding that five-view results are worse than three-view on München points to visibility fusion, rather than depth-range prediction, as the next bottleneck for aerial MVS, a direction the paper itself flags when discussing its lightweight visibility-aware fusion.
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 ADR-MVS, an aerial multi-view stereo network that integrates monocular depth and normal priors into three components: a depth range predictor (DRP) that generates per-pixel adaptive depth ranges for cascaded cost volumes, a normal-guided cost aggregation (NCA) that replaces standard 3D convolutions with geometry- and feature-aware aggregation, and a normal-guided depth refinement (NDR) module. The method is built on the GoMVS backbone and evaluated on the WHU, LuoJia-MVS, and München datasets. The authors report state-of-the-art MAE on all three benchmarks, with lower memory and runtime than several baselines, and support the design with ablations in Tables V-VII.

Significance. If the reported gains are robust, the paper makes a useful contribution by showing that frozen monocular geometric cues can be injected into cascaded MVS cost volume construction, cost aggregation, and refinement for aerial imagery. The ablations are thorough: they isolate the contributions of the DRP, NCA, and NDR, compare different monocular prior combinations (Table VII), and honestly discuss the 5-view degradation on München. The main weakness is that the state-of-the-art claim rests on very small margins over the closest aerial competitors, with no error bars or multiple-seed evaluation, and the computational complexity claim is presented in a way that excludes the monocular-prior inference cost.

major comments (3)
  1. [Tables II and III; Section IV-C] The central claim of state-of-the-art accuracy rests on margins of 0.1-0.4 cm over the closest aerial competitor (WHU-3: 9.4 vs 9.5 cm for SDL-MVS; LuoJia-3: 8.2 vs 8.6 cm) and 0.0015 m over EG-MVSNet on München with WHU-5 training (0.4812 vs 0.4827 m). These differences are within typical seed-to-seed variation for deep MVS training, and the paper reports no error bars, no multiple runs, and no significance tests. In addition, the SDL-MVS numbers are cited from EG-MVSNet rather than obtained by re-running the method. Please report mean±std over at least three independent training runs (or an equivalent paired comparison) and, where possible, re-run the competing aerial methods under the same protocol before claiming state-of-the-art status.
  2. [Section II.A, Table IV, Table VII, Conclusion] The efficiency claim is contradicted by the reported end-to-end numbers. Table IV lists 2836 MB for 'Ours' but excludes the monocular-prior inference step; Table VII shows that the prior step adds 625-1976 MB and up to 0.534 s, so the total memory for ADR-MVS is 3461-4812 MB versus 2891 MB for the GoMVS baseline, and the total runtime is not consistently lower. The statements that the method works 'without additional memory overhead' (Section II.A) and 'without incurring additional computational costs' (Conclusion) are therefore misleading unless they refer only to the matching pipeline. Please report end-to-end costs or qualify the claim explicitly.
  3. [Section III-A with Eq. (11)] The sentence in Section III-A states that 'the predicted depth range must be at least twice the monocular depth deviation,' but the residual loss in Eq. (11) does not enforce this. The loss penalizes (i) ground-truth depths falling outside the predicted range and (ii) deviations of R_pred from R_gt = |D_gt - D_m| (or from the previous-stage residual). That combination encourages R_pred ≈ R_gt, not a factor-of-two margin. The statement and the equation need to be reconciled, because as written the training objective does not match the described design.
minor comments (5)
  1. [Abstract and Introduction] There is a typo in the Introduction: 'we deigne a Normal-Guided Depth Refinement' should read 'we design a Normal-Guided Depth Refinement.'
  2. [Section III-D, Eq. (11)] The notation in Eq. (11), e.g., '|D_m ± R_k_pred, D_gt|2', is not standard and is difficult to parse. Please rewrite the loss terms with explicit indicator functions or boundary-margin notation, and define all operations (e.g., squared L2, absolute value) unambiguously.
  3. [Table II note] The note 'Some results are obtained from EG-MVSNet' is too vague. Please specify which rows and which metric values were taken from EG-MVSNet, and ideally re-run those methods with the same training protocol.
  4. [Section IV-B and Table VII note] For Omnidata normal preprocessing, the paper states that memory and runtime were not recorded for patch-based high-resolution inference. This makes the total-cost comparison in Table VII incomplete for the 'Omni' rows; please either include the preprocessing cost or clearly mark it as excluded in the table.
  5. [Section III-A] The claim that DRP is 'the first attempt to leverage large-scale monocular geometric models for aerial MVS' is difficult to verify and may be overstated, since prior works have used monocular depth or normal priors in close-range MVS. Consider softening the contribution claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the SOTA claim rests on held-out test evaluation, and the depth-range supervision is a training loss rather than a renamed prediction.

full rationale

The paper's central claim, that ADR-MVS achieves state-of-the-art MAE on WHU, LuoJia-MVS, and München, is supported by independent held-out evaluation against ground-truth depth maps. The monocular depth and normal priors (DepthAnythingV2, Omnidata, and ablations with DepthPro/Lotus) are externally pretrained models, not outputs of the proposed network, so they do not smuggle the final result in by construction. The Depth Range Predictor's range maps are supervised with a residual loss that penalizes ranges failing to contain GT depth and overly large ranges (Eq. 11); this is a legitimate training objective for an intermediate representation, not a circular definition of the final depth estimate, because the final depth is still obtained by feature matching, cost aggregation, and depth regression (Eq. 8) and is evaluated directly against GT on test splits. Ablations (Table V and Table VI) show that removing DRP, NCA, or NDR degrades MAE, which evidences that the proposed components contribute beyond simply fitting the training loss. The paper contains no load-bearing self-citation chains: the GoMVS baseline and monocular priors are external prior work, and the comparison numbers for some competitors are quoted from EG-MVSNet rather than re-run, which is a reproducibility concern but not a circularity. The absence of error bars or significance tests affects the strength of the SOTA claim, but it is not a derivation-reduction issue. Overall, the derivation chain from adaptive range prediction to final depth estimation is not equivalent to its inputs by construction.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

No new physical entities are introduced. The method relies on standard homography warping, cascaded cost volumes, and learned modules. The listed axioms are the main domain assumptions that are load-bearing for the depth range predictor and the cost aggregation.

free parameters (2)
  • Depth candidate counts per stage = 48, 32, 8
    Standard choices from GoMVS/cascaded MVS; the paper fixes them to keep memory comparable to baseline, so results may depend on this choice.
  • Monocular depth scaling rule = unspecified
    Section III-A: monocular depth is 'directly scaled' to the predefined range, but the scaling formula is not given, which is both a free choice and a reproducibility gap.
assumptions (4)
  • domain assumption Monocular depth and normal priors from pretrained models provide reliable and unbiased geometric cues for aerial scenes.
    The DRP's range prediction is built on the discrepancy between depth-derived normals and monocular normals; if monocular normals are biased, ranges fail. Introduced in Section III-A.
  • domain assumption Depth-to-normal conversion via the closed-form of [20] is sufficiently accurate.
    This conversion feeds the DRP; errors propagate to range maps. Section III-A.
  • domain assumption Pixels within the deformable neighborhood lie on approximately the same plane (Eq. 6).
    NCA aligns depth candidates using the plane equation; violated at depth edges. Section III-B.
  • domain assumption The predefined depth ranges from SfM are acceptable priors for the baseline and later stages.
    The loss and ablation compare against predefined ranges; if the SfM ranges were grossly wrong, the learning target would be ill-posed. Section IV-D.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Aerial Multi-View Stereo via Adaptive Depth Range Inference and Normal Cues." pith.science (2026). https://pith.science/paper/QU7XNFAG

@misc{pith2026250605655,
  author       = {Pith},
  title        = {Pith review of: Aerial Multi-View Stereo via Adaptive Depth Range Inference and Normal Cues},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QU7XNFAG}},
  note         = {Machine review of arXiv:2506.05655}
}
read the original abstract

Three-dimensional digital urban reconstruction from multi-view aerial images is a critical application where deep multi-view stereo (MVS) methods outperform traditional techniques. However, existing methods commonly overlook the key differences between aerial and close-range settings, such as varying depth ranges along epipolar lines and insensitive feature-matching associated with low-detailed aerial images. To address these issues, we propose an Adaptive Depth Range MVS (ADR-MVS), which integrates monocular geometric cues to improve multi-view depth estimation accuracy. The key component of ADR-MVS is the depth range predictor, which generates adaptive range maps from depth and normal estimates using cross-attention discrepancy learning. In the first stage, the range map derived from monocular cues breaks through predefined depth boundaries, improving feature-matching discriminability and mitigating convergence to local optima. In later stages, the inferred range maps are progressively narrowed, ultimately aligning with the cascaded MVS framework for precise depth regression. Moreover, a normal-guided cost aggregation operation is specially devised for aerial stereo images to improve geometric awareness within the cost volume. Finally, we introduce a normal-guided depth refinement module that surpasses existing RGB-guided techniques. Experimental results demonstrate that ADR-MVS achieves state-of-the-art performance on the WHU, LuoJia-MVS, and M\"unchen datasets, while exhibits superior computational complexity.

Figures

Figures reproduced from arXiv: 2506.05655 by the authors.

Figure 1
Figure 1. Feature matching scores for close-range vs. aerial image pairs. For a given pixel in the reference image (a-b), the initial depth ranges projected onto [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The overall framework of our proposed ADR-MVS. The network consists of three primary components: depth range map predictor, matching-based [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Illustration of Discrepancy Perception Module. FN denotes the [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Visualization comparison of depth estimation performance across different land types from the WHU (a) and LuoJia-MVS (b) datasets. The first and [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Visualization comparison of depth estimation performance on M [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Visualization of the intermediate results of two DRP designs. (b-c): The inputs of the MFE module. (d): Partial channels of the discrepancy feature [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Visualization of predicted range maps across multiple stages. (a) and (b) illustrate examples from the WHU dataset [ [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Qualitative results of ADR-MVS with different monocular priors on LuoJia-MVS 109 [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: Visualization of Point Cloud Reconstructions. Images with orange borders represent enlarged sections from the left scenes, while images with green borders depict views from different perspectives. V. CONCLUSION In this study, we presented a novel MVS method, ADR￾MVS, d…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 41 canonical work pages

  1. [1]

    Patchmatch stereo-stereo matching with slanted support windows

    M. Bleyer, C. Rhemann, and C. Rother, “Patchmatch stereo-stereo matching with slanted support windows.” inBritish Machine Vision Conference (BMVC), vol. 11, 2011, pp. 1–11

    work page 2011

  2. [2]

    Stereo processing by semiglobal matching and mutual information,

    H. Hirschmuller, “Stereo processing by semiglobal matching and mutual information,”IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), vol. 30, no. 2, pp. 328–341, 2007

    work page 2007

  3. [3]

    Dense stereo matching based on multiobjective fitness function—a genetic algorithm optimization approach for stereo correspondence,

    M. Manimala, G. Shirishkumar, J. Jyoti, and B. K. Mohan, “Dense stereo matching based on multiobjective fitness function—a genetic algorithm optimization approach for stereo correspondence,”IEEE Transactions Geoscience and Remote Sensing, vol. 57, no. 6, pp. 3341–3353, 2019. IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING 13

    work page 2019

  4. [4]

    Mvsnet: Depth infer- ence for unstructured multi-view stereo,

    Y . Yao, Z. Luo, S. Li, T. Fang, and L. Quan., “Mvsnet: Depth infer- ence for unstructured multi-view stereo,” inEuropean Conference on Computer Vision (ECCV), 2018

    work page 2018

  5. [5]

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

    Y . Yao, Z. Luo, S. Li, T. Shen, T. Fang, and L. Quan, “Recurrent mvsnet for high-resolution multi-view stereo depth inference,” inIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019, pp. 5525–5534

    work page 2019

  6. [6]

    Patch- matchnet: Learned multi-view patchmatch stereo,

    F. Wang, S. Galliani, C. V ogel, P. Speciale, and M. Pollefeys, “Patch- matchnet: Learned multi-view patchmatch stereo,” inIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020

    work page 2020

  7. [7]

    Curvature-guided dynamic scale net- works for multi-view stereo,

    G. K. Truong, S. Song, and S. Jo., “Curvature-guided dynamic scale net- works for multi-view stereo,” inInternational Conference on Learning Representations (ICLR), 2021

    work page 2021

  8. [8]

    Large- scale data for multiple-view stereopsis,

    A. Henrik, R. R. Jensen, G. V ogiatzis, E. Tola, and A. B. Dahl., “Large- scale data for multiple-view stereopsis,” inInternational Journal of Computer Vision, vol. 120, no. 2. Springer, 2016, pp. 153–168

    work page 2016

Show all 46 references
  1. [9]

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

    Y . Yao, Z. Luo, S. Li, J. Zhang, Y . Ren, L. Zhou, T. Fang, and L. Quan., “Blendedmvs: A large-scale dataset for generalized multi- view stereo networks,” inIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 1790–1799

    2020

  2. [10]

    Tanks and temples: Benchmarking large-scale scene reconstruction,

    K. Arno, J. Park, Q. Zhou, and V . Koltun., “Tanks and temples: Benchmarking large-scale scene reconstruction,”ACM Transactions on Graphics, vol. 36, no. 4, pp. 1–13, 2017

    2017

  3. [11]

    A hierarchical de- formable deep neural network and an aerial image benchmark dataset for surface multiview stereo reconstruction,

    J. Li, H. Xin, Y . Feng, Z. Ji, S. Zhang, and D. Wen., “A hierarchical de- formable deep neural network and an aerial image benchmark dataset for surface multiview stereo reconstruction,”IEEE Transactions Geoscience and Remote Sensing, vol. 61, pp. 1–12, 2023

    2023

  4. [12]

    A novel recurrent encoder-decoder structure for large- scale multi-view stereo reconstruction from an open aerial dataset,

    J. Liu and S. Ji., “A novel recurrent encoder-decoder structure for large- scale multi-view stereo reconstruction from an open aerial dataset,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 6050–6059

    2020

  5. [13]

    Edge aware depth inference for large-scale aerial building multi-view stereo,

    S. Zhang, Z. Wei, W. Xu, L. Zhang, Y . Wang, J. Zhang, and J. Liu., “Edge aware depth inference for large-scale aerial building multi-view stereo,”Journal of Photogrammetry and Remote Sensing, vol. 207, pp. 27–42, 2024

    2024

  6. [14]

    Bidirectional guided attention network for 3-d semantic detection of remote sensing images,

    Z. Rao, M. He, Z. Zhu, Y . Dai, and R. He., “Bidirectional guided attention network for 3-d semantic detection of remote sensing images,” IEEE Transactions Geoscience and Remote Sensing, vol. 59, no. 7, pp. 6138–6153, 2020

    2020

  7. [15]

    Gomvs: Geometrically consistent cost aggregation for multi-view stereo,

    J. Wu, R. Li, H. Xu, W. Zhao, Y . Zhu, J. Sun, and Y . Zhang, “Gomvs: Geometrically consistent cost aggregation for multi-view stereo,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 20 207–20 216

    2024

  8. [16]

    Depth anything v2,

    L. Yang, B. Kang, Z. Huang, Z. Zhao, X. Xu, J. Feng, and H. Zhao., “Depth anything v2,”arXiv preprint arXiv:2406.09414, 2024

    2024 arXiv

  9. [17]

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

    A. Bochkovskii, A. Delaunoy, H. Germain, M. Santos, Y . Zhou, S. R. Richter, and V . Koltun, “Depth pro: Sharp monocular metric depth in less than a second,”arXiv preprint arXiv:2410.02073, 2024

    2024 arXiv

  10. [18]

    Lotus: Diffusion-based visual foundation model for high-quality dense prediction,

    J. He, H. Li, W. Yin, Y . Liang, L. Li, K. Zhou, H. Liu, B. Liu, and Y .-C. Chen, “Lotus: Diffusion-based visual foundation model for high-quality dense prediction,”arXiv preprint arXiv:2409.18124, 2024

    2024 arXiv

  11. [19]

    Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans,

    A. Eftekhar, A. Sax, J. Malik, and A. Zamir, “Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans,” inIEEE International Conference on Computer Vision (ICCV), 2021, pp. 10 786–10 796

    2021

  12. [20]

    Geonet: Geometric neural network for joint depth and surface normal estimation,

    X. Qi, R. Liao, Z. Liu, R. Urtasun, and J. Jia, “Geonet: Geometric neural network for joint depth and surface normal estimation,” inIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018, pp. 283–291

    2018

  13. [21]

    A general deep learning based framework for 3d reconstruction from multi-view stereo satellite images,

    J. Gao, J. Liu, and S. Ji, “A general deep learning based framework for 3d reconstruction from multi-view stereo satellite images,”Journal of Photogrammetry and Remote Sensing, vol. 195, pp. 446–461, 2023

    2023

  14. [22]

    Sdl- mvs: View space and depth deformable learning paradigm for multi-view stereo reconstruction in remote sensing,

    Y . Mao, H. Bi, L. Xu, K. Chen, Z. Wang, X. Sun, and K. Fu., “Sdl- mvs: View space and depth deformable learning paradigm for multi-view stereo reconstruction in remote sensing,”IEEE Transactions Geoscience and Remote Sensing, 2024

    2024

  15. [23]

    A multi-task network for multi- view stereo reconstruction: When semantic consistency based clustering meets depth estimation optimization,

    H. Xin, S. Zhang, J. Li, and L. Wang., “A multi-task network for multi- view stereo reconstruction: When semantic consistency based clustering meets depth estimation optimization,”IEEE Transactions Geoscience and Remote Sensing, 2024

    2024

  16. [24]

    Deep learning based multi-view stereo matching and 3d scene reconstruction from oblique aerial images,

    J. Liu, J. Gao, S. Ji, C. Zeng, S. Zhang, and J. Gong, “Deep learning based multi-view stereo matching and 3d scene reconstruction from oblique aerial images,”Journal of Photogrammetry and Remote Sensing, vol. 204, pp. 42–60, 2023

    2023

  17. [25]

    Surface depth estimation from multi-view stereo satellite images with distribution contrast network,

    Z. Chen, W. Li, Z. Cui, and Y . Zhang, “Surface depth estimation from multi-view stereo satellite images with distribution contrast network,” Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2024

    2024

  18. [26]

    Sd-mvs: Segmentation- driven deformation multi-view stereo with spherical refinement and em optimization,

    Z. Yuan, J. Cao, Z. Li, H. Jiang, and Z. Wang, “Sd-mvs: Segmentation- driven deformation multi-view stereo with spherical refinement and em optimization,” inAAAI Conference on Artificial Intelligence, vol. 38, no. 7, 2024, pp. 6871–6880

    2024

  19. [27]

    Multi-view depth estimation by fusing single-view depth probability with multi-view geometry,

    G. Bae, I. Budvytis, and R. Cipolla, “Multi-view depth estimation by fusing single-view depth probability with multi-view geometry,” inIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022, pp. 2842–2851

    2022

  20. [28]

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

    L. Bartolomei, F. Tosi, M. Poggi, and S. Mattoccia, “Stereo anywhere: Robust zero-shot deep stereo matching even where either stereo or mono fail,”arXiv preprint arXiv:2412.04472, 2024

    2024 arXiv

  21. [29]

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

    Z. Chen, X. Ye, W. Yang, Z. Xu, X. Tan, Z. Zou, E. Ding, X. Zhang, and L. Huang, “Revealing the reciprocal relations between self-supervised stereo and monocular depth estimation,” inIEEE International Confer- ence on Computer Vision (ICCV), 2021, pp. 15 529–15 538

    2021

  22. [30]

    Learning stereo from single images,

    J. Watson, O. M. Aodha, D. Turmukhambetov, G. J. Brostow, and M. Fir- man, “Learning stereo from single images,” inEuropean Conference on Computer Vision (ECCV), 2020, pp. 722–740

    2020

  23. [31]

    Nr-mvsnet: Learning multi-view stereo based on normal consistency and depth refinement,

    J. Li, Z. Lu, Y . Wang, J. Xiao, and Y . Wang, “Nr-mvsnet: Learning multi-view stereo based on normal consistency and depth refinement,” IEEE Transactions on Image Processing, 2023

    2023

  24. [32]

    Enforcing geometric constraints of virtual normal for depth prediction,

    W. Yin, Y . Liu, C. Shen, and Y . Yan, “Enforcing geometric constraints of virtual normal for depth prediction,” inIEEE International Conference on Computer Vision (ICCV), 2019, pp. 5684–5693

    2019

  25. [33]

    Normal assisted stereo depth estimation,

    U. Kusupati, S. Cheng, R. Chen, and H. Su, “Normal assisted stereo depth estimation,” inIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 2189–2199

    2020

  26. [34]

    Occlusion-aware depth estimation with adaptive normal constraints,

    X. Long, L. Liu, C. Theobalt, and W. Wang, “Occlusion-aware depth estimation with adaptive normal constraints,” inEuropean Conference on Computer Vision (ECCV). Springer, 2020, pp. 640–657

    2020

  27. [35]

    Normal assisted pixel-visibility learning with cost aggregation for multiview stereo,

    W. Tong, X. Guan, J. Kang, P. Z. Sun, R. Law, P. Ghamisi, and E. Q. Wu, “Normal assisted pixel-visibility learning with cost aggregation for multiview stereo,”IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 12, pp. 24 686–24 697, 2022

    2022

  28. [36]

    Segment anything,

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Lo, P. Doll´ar, and R. Girshick, “Segment anything,”arXiv preprint arXiv:2304.02643, 2023

    2023 arXiv

  29. [37]

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

    X. Gu, Z. Fan, S. Zhu, Z. Dai, F. Tan, and P. Tan., “Cascade cost volume for high-resolution multi-view stereo and stereo matching,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020

    2020

  30. [38]

    Deep stereo using adaptive thin volume representation with uncertainty awareness,

    S. Cheng, Z. Xu, S. Zhu, Z. Li, L. E. Li, R. Ramamoorthi, and H. Su, “Deep stereo using adaptive thin volume representation with uncertainty awareness,”IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019

    2019

  31. [39]

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

    R. Peng, R. Wang, Z. Wang, Y . Lai, and R. Wang, “Rethinking depth estimation for multi-view stereo: A unified representation,”IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 8645–8654, 2022

    2022

  32. [40]

    Vis-mvsnet: Visibility- aware multi-view stereo network,

    J. Zhang, S. Li, Z. Luo, T. Fang, and Y . Yao, “Vis-mvsnet: Visibility- aware multi-view stereo network,”International Journal of Computer Vision, vol. 131, no. 1, pp. 199–214, 2023

    2023

  33. [41]

    Cost volume pyramid based depth inference for multi-view stereo,

    J. Yang, W. Mao, J. M. Alvarez, and M. Liu, “Cost volume pyramid based depth inference for multi-view stereo,”IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020

    2020

  34. [42]

    The landscape of dense image matching algorithms,

    N. Haala, “The landscape of dense image matching algorithms,”Pho- togrammetric week, vol. 13, pp. 271–284, 2013

    2013

  35. [43]

    Pixel- wise view selection for unstructured multi-view stereo,

    J. L. Sch ¨onberger, E. Zheng, J.-M. Frahm, and M. Pollefeys., “Pixel- wise view selection for unstructured multi-view stereo,” inEuropean Conference on Computer Vision (ECCV), 2016, p. 501–518

    2016

  36. [44]

    Multi-view stereo in the deep learning era: A comprehensive review,

    X. Wang, C. Wang, B. Liu, X. Zhou, L. Zhang, J. Zheng, and X. Bai, “Multi-view stereo in the deep learning era: A comprehensive review,” Displays, vol. 70, p. 102102, 2021

    2021

  37. [45]

    Deep learning for multi-view stereo via plane sweep: A survey,

    Q. Zhu, C. Min, Z. Wei, Y . Chen, and G. Wang, “Deep learning for multi-view stereo via plane sweep: A survey,”arXiv preprint arXiv:2106.15328, 2021

    2021 arXiv

  38. [46]

    Variational methods for normal integration,

    Y . Qu´eau, J.-D. Durou, and J.-F. Aujol., “Variational methods for normal integration,”Journal of Mathematical Imaging and Vision, vol. 60, pp. 609–632, 2018

    2018

Pith tools

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