Pith. sign in

REVIEW 4 major objections 6 minor 70 references

M3Depth: Wavelet-Enhanced Depth Estimation on Mars via Mutual Boosting of Dual-Modal Data

T0 review · 4 major / 6 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 The paper claims that a stereo depth network combining Haar-wavelet low-frequency feature extraction with a depth-normal consistency loss and iterative mutual refinement achieves state-of-the-art depth estimates on a synthetic Mars…

desk verdict Useful engineering contribution with a real reproducibility wrinkle: the headline numbers in Table III don't match the stated default configuration, and the abstract's 16% claim overstates the table. read the letter →

arxiv 2505.14159 v2 pith:Q7W2RKCM submitted 2025-05-20 cs.CV cs.RO

Junjie Li , Jiawei Wang , Miyu Li , Yu Liu , Yumei Wang , Haitao Xu This is my paper · ORCID
classification cs.CVcs.RO
keywords stereodepthestimationMarsroverperceptionwavelettransformlow-frequencyfeaturessurfacenormalconstraintiterativerefinementSimMars6K
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

Mars terrain is dominated by smooth, texture-poor surfaces that carry mostly low-frequency visual information, which makes learning-based stereo depth estimation unreliable. M3Depth is a stereo network built to counter both problems: its feature extractor uses a Haar-wavelet decomposition to retain and amplify low-frequency components while enlarging the receptive field, and its training objective includes a consistency loss that ties predicted depth gradients to predicted surface normals, followed by a ConvGRU refinement loop that iteratively boosts both outputs. The paper shows on the SimMars6K synthetic benchmark that this combination lowers Abs Rel to 0.089 and RMSE to 0.314, improving on the previous best method by about 14%, and it presents qualitative results on Zhurong rover imagery suggesting the model generalizes beyond simulation. If these results hold, rover navigation and obstacle avoidance on future Mars missions could get denser and more accurate 3D terrain perception in exactly the low-texture environments where current methods struggle.

What carries the argument

The load-bearing mechanism is a wavelet-enhanced convolutional kernel built from the 2-D Haar transform, applied recursively to low-frequency components and recombined through inverse wavelet reconstruction with an adaptive low-frequency amplification coefficient; this yields an exponentially growing receptive field without larger kernels. Around it, the model adds a consistency constraint that equates depth gradients derived from the predicted surface normals via the pinhole model with Sobel-estimated depth gradients, imposed as a Huber-norm loss. Finally, a ConvGRU-based iterative refinement module computes residual updates alternately for depth and surface normal maps, so each modality is repeatedly re-estimated with the other as context.

What would settle it

Take a set of real Zhurong or Perseverance stereo pairs, reconstruct dense ground-truth depth through photogrammetric 3D reconstruction, and compute Abs Rel and RMSE for M3Depth and CRE-Stereo on those images; if M3Depth no longer beats CRE-Stereo by a similar margin, the paper's generalization claim is refuted.

Watch

Extended reading notes

Core claim

The central claim is that wavelet-enhanced feature extraction plus depth-normal mutual boosting yields the current best learned stereo depth estimates for Mars-like terrain. On the paper's own terms, the discovery is empirical: jointly exploiting the low-frequency structure of Martian imagery and the geometric consistency between depth and surface normals moves stereo depth accuracy on SimMars6K from CRE-Stereo's Abs Rel 0.103 and RMSE 0.367 to M3Depth's 0.089 and 0.314, with corresponding gains in threshold accuracy and smoother, more continuous depth maps on real Zhurong images.

Load-bearing premise

The reported numbers come entirely from a synthetic dataset, and the real-Mars evidence is qualitative, so the central claim assumes that SimMars6K's low-texture appearance and NaTeCam camera model faithfully represent the conditions a rover will actually encounter.

Editorial extensions

If this is right

  • M3Depth's reported gains on SimMars6K (Abs Rel 0.089, RMSE 0.314) imply that wavelet low-frequency feature extraction combined with depth-normal consistency loss can be added to existing stereo pipelines to produce denser, more accurate depth on texture-poor terrain.
  • The depth-normal consistency loss is a self-contained geometric constraint that needs no external priors or large pre-trained models, which makes the approach feasible for rover onboard computation.
  • The iterative mutual-boosting refinement improves both depth and surface normal predictions in the ablations, so downstream tasks such as rock detection and terrain reconstruction could share the same improved geometric outputs.
  • The qualitative generalization to Zhurong imagery suggests that, with modest domain adaptation, the model could be deployed on future Mars rover missions.

Reading between the lines

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

  • The wavelet low-frequency enhancement is likely to transfer to other visually homogeneous environments, such as lunar plains, polar ice fields, or terrestrial deserts, since the frequency analysis motivation is not Mars-specific.
  • A cleaner attribution test would compare M3Depth against a variant using dilated convolutions with a matched receptive field; if the gain disappears, the benefit may come from the larger field rather than the frequency selectivity.
  • The depth-normal consistency loss could support self-supervised training on unlabeled real rover imagery, using the geometric relation between depth and normals as a training signal in place of the depth ground truth that real Zhurong images lack.
  • Because the paper itself observes surface normals are easier to predict than depth, the mutual-boosting refinement module might also improve monocular depth estimation, where no second view is available.
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 / 6 minor

Summary. The paper proposes M3Depth, an end-to-end stereo depth estimation model tailored to Mars rover imagery. The contributions are: (i) a wavelet-enhanced feature extractor (WEFE) that recursively decomposes the low-frequency component with Haar wavelets and aggregates multi-scale features; (ii) a consistency loss that couples depth gradients and surface normal predictions under a pinhole-camera model; and (iii) an iterative refinement module (IRM) that alternately updates depth and surface normal maps through ConvGRU blocks. The model is evaluated on the synthetic SimMars6K dataset and qualitatively on real Zhurong rover images, reporting Abs Rel 0.089, Sq Rel 0.058, RMSE 0.314, log10 0.038, and d1 0.905, which the paper claims are the best among six compared methods. Ablations support the individual contributions of WEFE, IRM, and the consistency loss, and the authors release a dataset with surface-normal labels.

Significance. If the reported results are reproducible at the stated configuration, M3Depth is a meaningful advance for learning-based depth estimation in low-texture planetary environments. The wavelet-based low-frequency emphasis is well motivated by spectral analysis of Martian images, and the joint depth-normal refinement with consistency loss is clearly described and ablated. The release of a surface-normal augmented benchmark dataset is a tangible resource for the community. However, the quantitative case is currently undermined by a mismatch between the main results table and the appendix tuning tables, and by an abstract-level improvement claim that no single reported metric supports. These issues are fixable but need to be resolved before the performance claims can be accepted.

major comments (4)
  1. [Section V-B-4 / Table III / Appendix B] The headline numbers in Table III (Abs Rel 0.089, Sq Rel 0.058, RMSE 0.314, log10 0.038, d1 0.905) do not correspond to the stated default configuration of 5 IRM iterations and WEFE level 3. Appendix B (Table IX row 5 and Table X row 3) reports 0.084, 0.059, 0.308, and 0.912 for that configuration. Table III's Abs Rel/RMSE/d1 match the 3-iteration row of Table IX, but Sq Rel differs from that row as well (0.058 vs 0.065). Because no code or checkpoints are released, the configuration that actually produced the main results cannot be determined from the paper. Please reconcile these numbers, state the exact configuration used for each main table, and quote the results for the stated default configuration consistently throughout.
  2. [Abstract / Section V-C-1] The abstract's claim of a '16% improvement in depth estimation accuracy compared to other state-of-the-art methods' is not supported by Table III. Over the next-best method CRE-Stereo, the relative improvements are 13.6% for Abs Rel, 14.5% for RMSE, 7.9% for Sq Rel, 19.1% for log10, and 2.0% for d1. No single metric in Table III yields 16%, and no aggregate metric is defined. Please either define the aggregate metric used for the 16% figure or revise the abstract to match the per-metric improvements reported in the table.
  3. [Section V-C-2 / Section V-D] The paper claims 'strong applicability in real-world Martian scenarios' and 'strong generalization ability' based on qualitative depth maps from Zhurong imagery (Fig. 15), which have no ground-truth depth and no quantitative error metrics. The only quantitative evaluation is on the synthetic SimMars6K dataset, and the paper's own Section V-D acknowledges that the model 'still has room for improvement when applied to real-world data from the Zhurong Rover'. Please temper the abstract and Section V-C-2 claims, or support them with a quantitative proxy on real data such as reprojection error, disparity consistency, or comparison with rover-measured distances.
  4. [Table III / Appendix B] All quantitative comparisons in Table III are single point estimates without standard deviations, confidence intervals, or multiple-seed repetitions. Given that the difference between the 3- and 5-iteration settings in Table IX (Abs Rel 0.089 vs 0.084; RMSE 0.314 vs 0.308) is comparable in magnitude to the claimed gains over CRE-Stereo, the absence of error bars leaves the statistical significance of the reported improvements unclear. Please report mean and variance over several random seeds, or at minimum state the random seed and any determinism guarantees in the evaluation protocol.
minor comments (6)
  1. [Eq. (13) / Appendix A Eq. (38)] In the denominator of Eq. (13), the term involving n_y contains 'y−uc' but should be 'v−vc' to be consistent with Eq. (14) and with the chain-rule derivation in Appendix A.
  2. [Section V-C-3 (Ablation Evaluation on Depth Estimation)] The text states that the full model Setup (d) achieves 'an RMSE of 0.311', but Table VI reports RMSE 0.314 for the same setup; please correct the text to match the table.
  3. [Section V-C-3 (Ablation on Consistency Constraints Loss Function)] In the loss-configuration listing, 'λs = 3' appears in the full-network loss definition; this should be 'λc = 3' to match the notation in Eq. (24).
  4. [Figure 17 caption] The caption reads 'Visualization of dataset examples and depth distribution characteristics', which appears to be copied from Figure 10; it should describe the IRM iteration-depth visualization.
  5. [Conclusion] The conclusion misspells the dataset as 'SimMar6k'; the correct name is SimMars6K.
  6. [Table III header] The threshold accuracy headers are mis-formatted: 'δ2 <1.252' and 'δ3 <1.253' should be 'δ2 < 1.25^2' and 'δ3 < 1.25^3' as defined in Eq. (31).

Circularity Check

0 steps flagged · score 0.0 of 10
Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The model relies on standard pinhole geometry, rectified stereo input, and the synthetic dataset as a proxy for Mars. No new physical entities are introduced. The main free choices are loss weights, refinement iterations, and wavelet levels, all tuned by validation.

free parameters (3)
  • Loss weights lambda_d, lambda_n, lambda_c = 2, 1, 3
    Hand-set hyperparameters in Eq. 24; the ablation in Table VIII shows performance varies with them, so they influence the reported gains.
  • IRM refinement iterations T = 5
    Selected by validation in Appendix B; the main table and appendix tables report different metrics at this setting.
  • Wavelet decomposition levels l = 3
    Selected by validation in Appendix B; affects receptive field size and final accuracy.
assumptions (4)
  • domain assumption Stereo image pairs are rectified, so corresponding points lie on the same scanline.
    Section III-B states this assumption; cost volume construction and disparity soft argmin depend on it.
  • domain assumption The scene can be represented as a smooth implicit surface F(X,Y,Z)=0 with normal (nx,ny,nz) and nz nonzero where the consistency loss is applied.
    Appendix A, Eqs. 36-39; the gradient identity divides by nz, so vertical surfaces and depth discontinuities violate it.
  • domain assumption SimMars6K synthetic imagery generated by OAISYS is a faithful proxy for real Martian low-texture terrain.
    Used as the sole quantitative benchmark; no real-world ground truth depth is evaluated.
  • domain assumption Ground-truth surface normals derived from depth using the methods in [27], [64] are reliable enough to supervise the normal branch.
    Section V-B-1; errors in these normals propagate into both normal supervision and the consistency loss.

how reviews work

0 comments
Cite this review

Pith. "Pith review of M3Depth: Wavelet-Enhanced Depth Estimation on Mars via Mutual Boosting of Dual-Modal Data." pith.science (2026). https://pith.science/paper/Q7W2RKCM

@misc{pith2026250514159,
  author       = {Pith},
  title        = {Pith review of: M3Depth: Wavelet-Enhanced Depth Estimation on Mars via Mutual Boosting of Dual-Modal Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Q7W2RKCM}},
  note         = {Machine review of arXiv:2505.14159}
}
read the original abstract

Depth estimation plays a great potential role in obstacle avoidance and navigation for further Mars exploration missions. Compared to traditional stereo matching, learning-based stereo depth estimation provides a data-driven approach to infer dense and precise depth maps from stereo image pairs. However, these methods always suffer performance degradation in environments with sparse textures and lacking geometric constraints, such as the unstructured terrain of Mars. To address these challenges, we propose M3Depth, a depth estimation model tailored for Mars rovers. Considering the sparse and smooth texture of Martian terrain, which is primarily composed of low-frequency features, our model incorporates a convolutional kernel based on wavelet transform that effectively captures low-frequency response and expands the receptive field. Additionally, we introduce a consistency loss that explicitly models the complementary relationship between depth map and surface normal map, utilizing the surface normal as a geometric constraint to enhance the accuracy of depth estimation. Besides, a pixel-wise refinement module with mutual boosting mechanism is designed to iteratively refine both depth and surface normal predictions. Experimental results on synthetic Mars datasets with depth annotations show that M3Depth achieves a 16% improvement in depth estimation accuracy compared to other state-of-the-art methods in depth estimation. Furthermore, the model demonstrates strong applicability in real-world Martian scenarios, offering a promising solution for future Mars exploration missions.

Figures

Figures reproduced from arXiv: 2505.14159 by the authors.

Figure 1
Figure 1. An illustration of China Zhurong Rover stereo camera. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 3
Figure 3. A comparison between a Martian image and a terrestrial image. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 2
Figure 2. Traditional stereo matching algorithm pipeline. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figures from the paper (12 more)
Figure 5
Figure 5. Figure 5: Overview of our M3Depth model. In our model, the left image is served as the reference image and the right image as the target image. The cost volumes are constructed based on the wavelet-enhanced feature extraction network. estimation in feature-sparse environments? B…
Figure 6
Figure 6. Figure 6: Overview of the surface normal prediction network. The outputs are summed and normalized to generate the final predicted surface normal map. [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Illustration of hierarchical decomposition using wavelet transform. [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Pinhole camera model and the scene implicit function. [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 10
Figure 10. Figure 10: Visualization of dataset examples and depth distribution character [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]
Figure 11
Figure 11. Figure 11: Comparative analysis of depth estimation performance across [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]
Figure 12
Figure 12. Figure 12: Qualitative comparison of depth estimation results on Martian images. The first column shows the original reference images (left image), while the [PITH_FULL_IMAGE:figures/full_fig_p014_12.png]
Figure 13
Figure 13. Figure 13: Qualitative depth estimation results on the SimMars6K dataset (Case-I). The red boxes highlight the regions to emphasize. [PITH_FULL_IMAGE:figures/full_fig_p015_13.png]
Figure 14
Figure 14. Figure 14: Qualitative depth estimation results on the SimMars6K dataset (Case-II). [PITH_FULL_IMAGE:figures/full_fig_p015_14.png]
Figure 15
Figure 15. Figure 15: Qualitative evaluation of depth estimation results on Zhurong Rover images, which demonstrates the generalization ability of our method to real-world [PITH_FULL_IMAGE:figures/full_fig_p016_15.png]
Figure 16
Figure 16. Figure 16: Ablation study visualization results. Comparison of depth estimation results under different ablation settings. [PITH_FULL_IMAGE:figures/full_fig_p017_16.png]
Figure 17
Figure 17. Figure 17: Visualization of dataset examples and depth distribution character [PITH_FULL_IMAGE:figures/full_fig_p022_17.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

70 extracted references · 53 canonical work pages

  1. [17]

    Wavelet convo- lutions for large receptive fields,

    S. E. Finder, R. Amoyal, E. Treister, and O. Freifeld, “Wavelet convo- lutions for large receptive fields,” inEuropean Conference on Computer Vision. Springer, 2025, pp. 363–380

    work page 2025

  2. [27]

    Normal assisted stereo depth estimation,

    U. Kusupati, S. Cheng, R. Chen, and H. Su, “Normal assisted stereo depth estimation,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 2189–2199

    work page 2020

  3. [1]

    Perception-aware autonomous mast motion planning for planetary exploration rovers,

    J. Strader, K. Otsu, and A.-a. Agha-mohammadi, “Perception-aware autonomous mast motion planning for planetary exploration rovers,” Journal of Field Robotics, vol. 37, no. 5, pp. 812–829, 2020

    work page 2020

  4. [2]

    Driving curiosity: Mars rover mobility trends during the first seven years,

    A. Rankin, M. Maimone, J. Biesiadecki, N. Patel, D. Levine, and O. Toupet, “Driving curiosity: Mars rover mobility trends during the first seven years,” in2020 IEEE Aerospace Conference. IEEE, 2020, pp. 1–19

    work page 2020

  5. [3]

    Slip estimation for mars rover zhurong based on data drive,

    T. Zhang, S. Peng, Y . Jia, H. Tian, J. Sun, and C. Yan, “Slip estimation for mars rover zhurong based on data drive,”Applied Sciences, vol. 12, no. 3, p. 1676, 2022

    work page 2022

  6. [4]

    Metric3d: Towards zero-shot metric 3d prediction from a sin- gle image,

    W. Yin, C. Zhang, H. Chen, Z. Cai, G. Yu, K. Wang, X. Chen, and C. Shen, “Metric3d: Towards zero-shot metric 3d prediction from a sin- gle image,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 9043–9053

    work page 2023

  7. [5]

    Self-supervised learning to guide scientifically relevant categorization of martian terrain images,

    T. Panambur, D. Chakraborty, M. Meyer, R. Milliken, E. Learned- Miller, and M. Parente, “Self-supervised learning to guide scientifically relevant categorization of martian terrain images,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 1322–1332

    work page 2022

  8. [6]

    3-d semantic terrain reconstruction of monocular close-up images of martian terrains,

    P. Tian, M. Yao, X. Xiao, B. Zheng, T. Cao, Y . Xi, H. Liu, and H. Cui, “3-d semantic terrain reconstruction of monocular close-up images of martian terrains,”IEEE Transactions on Geoscience and Remote Sensing, vol. 62, pp. 1–16, 2024

    work page 2024

Show all 70 references
  1. [7]

    Swindepth: Unsupervised depth estimation using monocular sequences via swin transformer and densely cascaded network,

    D. Shim and H. J. Kim, “Swindepth: Unsupervised depth estimation using monocular sequences via swin transformer and densely cascaded network,” in2023 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2023, pp. 4983–4990

    2023

  2. [8]

    M2cs: A multimodal and campus-scapes dataset for dynamic slam and moving object perception,

    H. Zhao, M. Yao, Y . Zhao, Y . Jiang, H. Zhang, X. Xiao, and K. Gao, “M2cs: A multimodal and campus-scapes dataset for dynamic slam and moving object perception,”Journal of Field Robotics, vol. 42, no. 3, pp. 787–805, 2025

    2025

  3. [9]

    Ai4mars: A dataset for terrain-aware autonomous driving on mars,

    R. M. Swan, D. Atha, H. A. Leopold, M. Gildner, S. Oij, C. Chiu, and M. Ono, “Ai4mars: A dataset for terrain-aware autonomous driving on mars,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 1982–1991

    2021

  4. [10]

    Light4mars: A lightweight transformer model for semantic segmentation on unstruc- tured environment like mars,

    Y . Xiong, X. Xiao, M. Yao, H. Cui, and Y . Fu, “Light4mars: A lightweight transformer model for semantic segmentation on unstruc- tured environment like mars,”ISPRS Journal of Photogrammetry and Remote Sensing, vol. 214, pp. 167–178, 2024

    2024

  5. [11]

    Automated rock detection from mars rover image via y-shape dual-task network with depth-aware spatial attention mechanism,

    C. Ma, Y . Li, J. Lv, Z. Xiao, W. Zhang, and L. Mo, “Automated rock detection from mars rover image via y-shape dual-task network with depth-aware spatial attention mechanism,”IEEE Transactions on Geoscience and Remote Sensing, 2024

    2024

  6. [12]

    Geomorphic contexts and science focus of the zhurong landing site on mars,

    J. Liu, C. Li, R. Zhang, W. Rao, X. Cui, Y . Geng, Y . Jia, H. Huang, X. Ren, W. Yanet al., “Geomorphic contexts and science focus of the zhurong landing site on mars,”Nature Astronomy, vol. 6, no. 1, pp. 65–71, 2022. 19

    2022

  7. [13]

    The mars 2020 perseverance rover mast camera zoom (mastcam-z) multispectral, stereoscopic imaging investigation,

    J. Bell, J. Maki, G. Mehall, M. Ravine, M. Caplinger, Z. Bailey, S. Brylow, J. Schaffner, K. Kinch, M. Madsenet al., “The mars 2020 perseverance rover mast camera zoom (mastcam-z) multispectral, stereoscopic imaging investigation,”Space science reviews, vol. 217, pp. 1–40, 2021

    2020

  8. [14]

    A hybrid attention semantic seg- mentation network for unstructured terrain on mars,

    H. Liu, M. Yao, X. Xiao, and H. Cui, “A hybrid attention semantic seg- mentation network for unstructured terrain on mars,”Acta Astronautica, vol. 204, pp. 492–499, 2023

    2023

  9. [15]

    Unsupervised denoising with implicit noise mapping for single martian multispectral image,

    J. Li, W. Lv, J. Wang, Y . Wang, and Y . Liu, “Unsupervised denoising with implicit noise mapping for single martian multispectral image,” in IGARSS 2024-2024 IEEE International Geoscience and Remote Sensing Symposium. IEEE, 2024, pp. 6120–6123

    2024

  10. [16]

    Dustnet: An unsupervised and noise-resistant network for martian dust storm change detection,

    M. Li, J. Li, Y . Wang, Y . Liu, and H. Xu, “Dustnet: An unsupervised and noise-resistant network for martian dust storm change detection,” IEEE Geoscience and Remote Sensing Letters, 2025

    2025

  11. [18]

    Can biases in imagenet models explain gen- eralization?

    P. Gavrikov and J. Keuper, “Can biases in imagenet models explain gen- eralization?” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 22 184–22 194

    2024

  12. [19]

    Cascaded recurrent networks with masked representation learning for stereo matching of high-resolution satellite images,

    Z. Rao, X. Li, B. Xiong, Y . Dai, Z. Shen, H. Li, and Y . Lou, “Cascaded recurrent networks with masked representation learning for stereo matching of high-resolution satellite images,”ISPRS Journal of Photogrammetry and Remote Sensing, vol. 218, pp. 151–165, 2024

    2024

  13. [20]

    Efficient deep learning for stereo matching,

    W. Luo, A. G. Schwing, and R. Urtasun, “Efficient deep learning for stereo matching,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 5695–5703

    2016

  14. [21]

    Stereo matching by training a convolutional neural network to compare image patches,

    J. ˇZbontar and Y . LeCun, “Stereo matching by training a convolutional neural network to compare image patches,”Journal of Machine Learning Research, vol. 17, no. 65, pp. 1–32, 2016

    2016

  15. [22]

    Revisiting stereo depth estimation from a sequence-to-sequence perspective with transformers,

    Z. Li, X. Liu, N. Drenkow, A. Ding, F. X. Creighton, R. H. Tay- lor, and M. Unberath, “Revisiting stereo depth estimation from a sequence-to-sequence perspective with transformers,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 6197–6206

    2021

  16. [23]

    H-net: Un- supervised attention-based stereo depth estimation leveraging epipolar geometry,

    B. Huang, J.-Q. Zheng, S. Giannarou, and D. S. Elson, “H-net: Un- supervised attention-based stereo depth estimation leveraging epipolar geometry,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 4460–4467

    2022

  17. [24]

    Deep learning-based incorporation of planar constraints for robust stereo depth estimation in autonomous vehicle applications,

    W. Chuah, R. Tennakoon, R. Hoseinnezhad, and A. Bab-Hadiashar, “Deep learning-based incorporation of planar constraints for robust stereo depth estimation in autonomous vehicle applications,”IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 7, pp. 6654–6665, 2021

    2021

  18. [25]

    Monocular depth estimation with guidance of surface normal map,

    H. Yan, S. Zhang, Y . Zhang, and L. Zhang, “Monocular depth estimation with guidance of surface normal map,”Neurocomputing, vol. 280, pp. 86–100, 2018

    2018

  19. [26]

    Rethinking inductive biases for surface normal estimation,

    G. Bae and A. J. Davison, “Rethinking inductive biases for surface normal estimation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 9535–9545

    2024

  20. [28]

    S5mars: Semi- supervised learning for mars semantic segmentation,

    J. Zhang, L. Lin, Z. Fan, W. Wang, and J. Liu, “S5mars: Semi- supervised learning for mars semantic segmentation,”IEEE Transactions on Geoscience and Remote Sensing, vol. 62, p. 3386756, 2024

    2024

  21. [29]

    Marsnet: Automated rock segmentation with transformers for tianwen-1 mission,

    W. Lv, L. Wei, D. Zheng, Y . Liu, and Y . Wang, “Marsnet: Automated rock segmentation with transformers for tianwen-1 mission,”IEEE Geoscience and Remote Sensing Letters, vol. 20, pp. 1–5, 2022

    2022

  22. [30]

    Rockformer: A u-shaped transformer network for martian rock segmentation,

    H. Liu, M. Yao, X. Xiao, and Y . Xiong, “Rockformer: A u-shaped transformer network for martian rock segmentation,”IEEE Transactions on Geoscience and Remote Sensing, vol. 61, pp. 1–16, 2023

    2023

  23. [31]

    Marsformer: Martian rock semantic segmentation with transformer,

    Y . Xiong, X. Xiao, M. Yao, H. Liu, H. Yang, and Y . Fu, “Marsformer: Martian rock semantic segmentation with transformer,”IEEE Transac- tions on Geoscience and Remote Sensing, vol. 61, pp. 1–12, 2023

    2023

  24. [32]

    Marsscapes and udaformer: A panorama dataset and a transformer-based unsupervised domain adaptation framework for martian terrain segmentation,

    H. Liu, M. Yao, X. Xiao, B. Zheng, and H. Cui, “Marsscapes and udaformer: A panorama dataset and a transformer-based unsupervised domain adaptation framework for martian terrain segmentation,”IEEE Transactions on Geoscience and Remote Sensing, vol. 62, pp. 1–17, 2023

    2023

  25. [33]

    Rover attitude and camera parameter: Rock measurements on mars surface based on rover attitude and camera parameter for tianwen-1 mission,

    D. Zheng, L. Wei, W. Lv, Y . Liu, and Y . Wang, “Rover attitude and camera parameter: Rock measurements on mars surface based on rover attitude and camera parameter for tianwen-1 mission,”Remote Sensing, vol. 15, no. 18, p. 4388, 2023

    2023

  26. [34]

    High-resolution and spatial-continuous 3d model reconstruction of martian surface by integrating multi-sensor data of zhurong rover,

    R. Liu, Y . Xu, and Q. Yang, “High-resolution and spatial-continuous 3d model reconstruction of martian surface by integrating multi-sensor data of zhurong rover,”IEEE Transactions on Geoscience and Remote Sensing, 2024

    2024

  27. [35]

    Fast cost-volume filtering for visual correspondence and beyond,

    A. Hosni, C. Rhemann, M. Bleyer, C. Rother, and M. Gelautz, “Fast cost-volume filtering for visual correspondence and beyond,”IEEE transactions on pattern analysis and machine intelligence, vol. 35, no. 2, pp. 504–511, 2012

    2012

  28. [36]

    Aanet: Adaptive aggregation network for efficient stereo matching,

    H. Xu and J. Zhang, “Aanet: Adaptive aggregation network for efficient stereo matching,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 1959–1968

    2020

  29. [37]

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

    F. Zhang, V . Prisacariu, R. Yang, and P. H. Torr, “Ga-net: Guided aggregation net for end-to-end stereo matching,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 185–194

    2019

  30. [38]

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

    H. Laga, L. V . Jospin, F. Boussaid, and M. Bennamoun, “A survey on deep learning techniques for stereo-based depth estimation,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 4, pp. 1738–1764, 2022

    2022

  31. [39]

    Unifying flow, stereo and depth estimation,

    H. Xu, J. Zhang, J. Cai, H. Rezatofighi, F. Yu, D. Tao, and A. Geiger, “Unifying flow, stereo and depth estimation,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 11, pp. 13 941– 13 958, 2023

    2023

  32. [40]

    Robust depth estimation based on parallax attention for aerial scene perception,

    W. Tong, M. Zhang, G. Zhu, X. Xu, and E. Q. Wu, “Robust depth estimation based on parallax attention for aerial scene perception,”IEEE Transactions on Industrial Informatics, 2024

    2024

  33. [41]

    Displets: Resolving stereo ambiguities using object knowledge,

    F. Guney and A. Geiger, “Displets: Resolving stereo ambiguities using object knowledge,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2015, pp. 4165–4175

    2015

  34. [42]

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

    A. Kendall, H. Martirosyan, S. Dasgupta, P. Henry, R. Kennedy, A. Bachrach, and A. Bry, “End-to-end learning of geometry and context for deep stereo regression,” inProceedings of the IEEE international conference on computer vision, 2017, pp. 66–75

    2017

  35. [43]

    Improved stereo matching with constant highway networks and reflective confidence learning,

    A. Shaked and L. Wolf, “Improved stereo matching with constant highway networks and reflective confidence learning,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 4641–4650

    2017

  36. [44]

    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,” inProceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 5684–5693

    2019

  37. [45]

    Depthfm: Fast monocular depth estimation with flow matching,

    M. Gui, J. S. Fischer, U. Prestel, P. Ma, D. Kotovenko, O. Grebenkova, S. A. Baumann, V . T. Hu, and B. Ommer, “Depthfm: Fast monocular depth estimation with flow matching,”arXiv preprint arXiv:2403.13788, 2024

    2024 arXiv

  38. [46]

    Irondepth: Iterative refinement of single-view depth using surface normal and its uncertainty,

    G. Bae, I. Budvytis, and R. Cipolla, “Irondepth: Iterative refinement of single-view depth using surface normal and its uncertainty,”arXiv preprint arXiv:2210.03676, 2022

    2022 arXiv

  39. [47]

    Nddepth: Normal-distance assisted monocular depth estimation,

    S. Shao, Z. Pei, W. Chen, X. Wu, and Z. Li, “Nddepth: Normal-distance assisted monocular depth estimation,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 7931–7940

    2023

  40. [48]

    A kernel- based multi-featured rock modeling and detection framework for a mars rover,

    X. Xiao, M. Yao, H. Liu, J. Wang, L. Zhang, and Y . Fu, “A kernel- based multi-featured rock modeling and detection framework for a mars rover,”IEEE Transactions on Neural Networks and Learning Systems, vol. 34, no. 7, pp. 3335–3344, 2021

    2021

  41. [49]

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

    X. Wang, G. Xu, H. Jia, and X. Yang, “Selective-stereo: Adaptive frequency information selection for stereo matching,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 19 701–19 710

    2024

  42. [50]

    High- frequency stereo matching network,

    H. Zhao, H. Zhou, Y . Zhang, J. Chen, Y . Yang, and Y . Zhao, “High- frequency stereo matching network,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 1327– 1336

    2023

  43. [51]

    Stereo processing by semiglobal matching and mu- tual information,

    H. Hirschmuller, “Stereo processing by semiglobal matching and mu- tual information,”IEEE Transactions on pattern analysis and machine intelligence, vol. 30, no. 2, pp. 328–341, 2007

    2007

  44. [52]

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

    D. Scharstein and R. Szeliski, “A taxonomy and evaluation of dense two-frame stereo correspondence algorithms,”International journal of computer vision, vol. 47, pp. 7–42, 2002

    2002

  45. [53]

    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,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018, pp. 283–291

    2018

  46. [54]

    Rdrf-net: A pyramid architecture network with residual-based dynamic receptive fields for unsupervised depth estimation,

    Z. Ji, X. Song, H. Song, H. Yang, and X. Guo, “Rdrf-net: A pyramid architecture network with residual-based dynamic receptive fields for unsupervised depth estimation,”Neurocomputing, vol. 457, pp. 1–12, 2021

    2021

  47. [55]

    Wavelet-inspired sparse learning network for hyperspectral image change detection,

    C. Zhou, Z. He, J. Dong, and L. Zou, “Wavelet-inspired sparse learning network for hyperspectral image change detection,”IEEE Geoscience and Remote Sensing Letters, 2025. 20

    2025

  48. [56]

    Hyperspectral image classification using cnn-enhanced multi-level haar wavelet features fusion network,

    W. Guo, G. Xu, B. Liu, and Y . Wang, “Hyperspectral image classification using cnn-enhanced multi-level haar wavelet features fusion network,” IEEE Geoscience and Remote Sensing Letters, vol. 19, pp. 1–5, 2022

    2022

  49. [57]

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

    S. Im, H.-G. Jeon, S. Lin, and I. S. Kweon, “Dpsnet: End-to-end deep plane sweep stereo,”arXiv preprint arXiv:1905.00538, 2019

    1905 arXiv

  50. [58]

    Attention concatenation volume for accurate and efficient stereo matching,

    G. Xu, J. Cheng, P. Guo, and X. Yang, “Attention concatenation volume for accurate and efficient stereo matching,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 12 981–12 990

    2022

  51. [59]

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

    Z. Shen, Y . Dai, and Z. Rao, “Cfnet: Cascade and fused cost volume for robust stereo matching,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 13 906–13 915

    2021

  52. [60]

    Multi- level context ultra-aggregation for stereo matching,

    G. Nie, M. Cheng, Y . Liu, Z. Liang, D. Fan, Y . Liu, and Y . Wang, “Multi- level context ultra-aggregation for stereo matching,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 3283–3291

    2019

  53. [61]

    Semantic stereo matching with pyramid cost volumes,

    Z. Wu, X. Wu, X. Zhang, S. Wang, and L. Ju, “Semantic stereo matching with pyramid cost volumes,” inProceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 7484–7493

    2019

  54. [62]

    Raft-stereo: Multilevel recurrent field transforms for stereo matching,

    L. Lipson, Z. Teed, and J. Deng, “Raft-stereo: Multilevel recurrent field transforms for stereo matching,” in2021 International Conference on 3D Vision (3DV). IEEE, 2021, pp. 218–227

    2021

  55. [63]

    Pan-location mapping and localization for the in-situ science exploration of zhurong mars rover,

    X. Zeng, J. Liu, X. Ren, W. Yan, Q. Fu, X. Gao, W. Chen, W. Zuo, and C. Li, “Pan-location mapping and localization for the in-situ science exploration of zhurong mars rover,”Science China Information Sciences, vol. 65, no. 7, p. 172201, 2022

    2022

  56. [64]

    Data-driven 3d primitives for single image understanding,

    D. F. Fouhey, A. Gupta, and M. Hebert, “Data-driven 3d primitives for single image understanding,” inProceedings of the IEEE International Conference on Computer Vision, 2013, pp. 3392–3399

    2013

  57. [65]

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

    V . Tankovich, C. Hane, Y . Zhang, A. Kowdle, S. Fanello, and S. Bouaziz, “Hitnet: Hierarchical iterative tile refinement network for real-time stereo matching,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 14 362–14 372

    2021

  58. [66]

    Practical stereo matching via cascaded recurrent network with adaptive correlation,

    J. Li, P. Wang, P. Xiong, T. Cai, Z. Yan, L. Yang, J. Liu, H. Fan, and S. Liu, “Practical stereo matching via cascaded recurrent network with adaptive correlation,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 16 263–16 272

    2022

  59. [67]

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

    R. Peng, R. Wang, Z. Wang, Y . Lai, and R. Wang, “Rethinking depth es- timation for multi-view stereo: A unified representation,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 8645–8654

    2022

  60. [68]

    Deep surface normal estimation with hierarchical rgb-d fusion,

    J. Zeng, Y . Tong, Y . Huang, Q. Yan, W. Sun, J. Chen, and Y . Wang, “Deep surface normal estimation with hierarchical rgb-d fusion,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 6153–6162. Junjie Li(Graduate Student Member, I...

    2019

  61. [70]

    This loss is defined as the Huber norm of their deviation, as shown in Eq.(41)

    The consistency constraint is further formulated as a consistency loss function, enabling it to guide model training by aligning the two gradient estimates. This loss is defined as the Huber norm of their deviation, as shown in Eq.(41). Lc = ∂Z ∂u , ∂Z ∂v 1 − ∂Z ∂u , ∂Z ∂v 2 H...

  62. [2001]

    Currently she is an associate professor in Bei- jing University of Posts and Telecommunications

    From April 2011 to April 2012, she was working as a visiting scholar in Lehigh University, USA. Currently she is an associate professor in Bei- jing University of Posts and Telecommunications. Her research interests include multimedia signal pro- cessing, cross-layer design fo...

    2011

Pith tools

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