Pith. sign in

REVIEW 5 major objections 7 minor 30 references

EndoMUST: Monocular Depth Estimation for Robotic Endoscopy via End-to-end Multi-step Self-supervised Training

T0 review · 5 major / 7 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 The paper claims that dividing each end-to-end training epoch into three isolated steps — optical flow registration, multiscale intrinsic image decomposition, and rigid transformation alignment — produces state-of-the-art self-supervised…

desk verdict Solid empirical integration paper with a reproducible SOTA claim held back by an unstated scale-alignment procedure and test-set selection of a warm-up step. read the letter →

arxiv 2506.16017 v1 pith:BOV4Y3CW submitted 2025-06-19 cs.CV cs.RO

Liangjing Shao , Linxin Bai , Chenkang Du , Xinrong Chen This is my paper · ORCID
classification cs.CVcs.RO
keywords monoculardepthestimationself-supervisedlearningroboticendoscopyintrinsicimagedecompositionopticalflowparameter-efficientfine-tuningDV-LoRAzero-shotgeneralization
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

This paper tries to show that the way a self-supervised monocular-depth training epoch is sequenced matters as much as the modules inside it. The authors propose a three-step schedule — optical-flow registration first, then multiscale intrinsic image decomposition, then rigid transformation alignments using predicted depth, ego-motion, and camera intrinsics — so that each network is updated only when its own supervision signal is active. On the SCARED dataset the method reports the lowest errors among compared self-supervised approaches, and the same weights transfer to the Hamlyn dataset with the best zero-shot numbers. The practical stake is that accurate depth could make robot-assisted endoscopy safer and more autonomous without requiring ground-truth depth labels.

What carries the argument

The load-bearing mechanism is the staged epoch: Step I trains optical-flow and appearance-flow networks with a registration loss against a visibility-masked target; Step II decomposes images at scales 1.0, 0.75, and 0.5 into reflectance and shading and supervises reconstruction; Step III aligns the source image, its intrinsic images, and the reflectance image to the target using predicted depth, ego-motion, and intrinsics, adding depth and appearance smoothness terms. The final component is DV-LoRA, a low-rank adaptation that modifies each transformer block's QKV and feed-forward linear layers with trainable diagonal scaling matrices around frozen weights, giving the depth network new capacity while adding only about 0.2 million trainable parameters.

What would settle it

Rerun the SCARED and Hamlyn evaluations with an explicitly specified scale alignment (for example, per-image median scaling versus a least-squares fit to ground truth) for every method; if EndoMUST's advantage over EndoDAC and MonoPCC shrinks or reverses under one consistent convention, the reported ranking depends on an unstated choice rather than on the three-step training.

Watch

Extended reading notes

Core claim

The central claim is that a three-step, end-to-end training strategy, built on top of a parameter-efficiently fine-tuned Depth Anything foundation model, reaches state-of-the-art self-supervised depth estimation for robotic endoscopy. In each epoch, optical flow and appearance flow networks are optimized alone; then an intrinsic-image-decomposition network is trained at three image scales; finally depth, ego-motion, camera-intrinsics, and appearance networks are trained together through photometric and illumination-free reflectance alignments. Reported numbers on SCARED are RelAbs 0.046, RelSq 0.313, RMSE 4.276, RMSElog 0.067, delta 0.984, beating the previous best compared method, and zero-shot Hamlyn results are RelAbs 0.145 and delta 0.807. The paper attributes the improvement to isolating constraints so that irrelevant information cannot mislead a given module, together with full-module DV-LoRA fine-tuning that keeps trainable parameters low.

Load-bearing premise

The comparison assumes that all predicted depth maps were aligned to ground truth by the same standard procedure, but the paper never describes how the scale-ambiguous predicted depth was scaled before computing RelAbs, RMSE, or delta.

Editorial extensions

If this is right

  • Self-supervised depth for endoscopy does not require ground truth, and with staged training it can reach sub-5% relative error on SCARED.
  • The same SCARED-trained weights generalize zero-shot to Hamlyn, over 90,000 frames, while keeping the best reported accuracy among compared methods.
  • Parameter-efficient fine-tuning lets a large foundation model adapt with only about 1.8M trainable parameters in the depth network, keeping inference above 100 fps on SCARED and above 200 fps on Hamlyn.
  • Ego-motion and camera-intrinsics estimates also improve under the third step, with focal-length relative error dropping below 1% in the reported sequences.
  • If the results hold, monocular endoscopic depth becomes a low-cost basis for 3D reconstruction, registration, and augmented-reality navigation during robot-assisted surgery.

Reading between the lines

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

  • The staged schedule may be transferable to other self-supervised geometric tasks where multiple constraints compete, such as monocular odometry in low-texture medical scenes; the paper does not test this.
  • The ablation tables suggest warm-up length of DV-LoRA is a sensitive hyperparameter, since 5000 warm-up steps degrade results, so part of the reported gain may be due to the tuned warm-up rather than the three-step ordering.
  • Because the paper omits the scale-alignment protocol for metrics, the precise numerical state-of-the-art claim is not independently reproducible until that protocol is specified; a reader should treat the ranking as tentative.
  • A direct way to strengthen the claim would be a per-sequence or per-video analysis showing that staging helps most in exactly the illumination-challenging frames the method targets; the paper reports aggregate numbers only.
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

5 major / 7 minor

Summary. The paper proposes EndoMUST, a self-supervised monocular depth and ego-motion estimation framework for robotic endoscopy. The main contribution is a training strategy that splits each epoch into three steps: optical flow registration, multiscale intrinsic image decomposition, and multiple transformation alignments, with only the task-relevant networks trainable at each step. The depth network is finetuned from a foundation model (Depth Anything) using DV-LoRA. The authors report state-of-the-art results on SCARED for self-supervised depth estimation and on Hamlyn for zero-shot depth estimation, along with ego-motion and camera-intrinsics results. The paper includes ablations over step order, step fusion, the multiscale decomposition, the QKV finetuning, and warm-up steps.

Significance. If the reported results hold, the paper makes a useful practical contribution: it shows that a staged, interference-avoiding training schedule combined with parameter-efficient finetuning of a foundation model improves self-supervised depth estimation in endoscopy. The strength of the paper is its clear and systematic experimental design: the ablations in Table V directly test the paper's central claim that the three-step strategy, rather than any single module, is responsible for the gains. The release of evaluation code is also a positive. However, the headline SOTA claim rests on an incompletely specified evaluation protocol (scale alignment), on hyperparameters that appear to be selected using the test set, and on small performance margins without uncertainty quantification. These issues currently prevent the claim from being verified from the manuscript alone.

major comments (5)
  1. [Section IV-C, Eqs. (13)-(16)] The paper never specifies how the scale-ambiguous predicted depth is aligned to the ground truth before computing RelAbs, RelSq, RMSE, RMSElog, and delta. Since self-supervised monocular depth is defined only up to an unknown scale, the reported numbers in Tables I and II require an alignment convention, e.g., median scaling per image, least-squares fitting, or per-sequence alignment. The margins over the best baselines are small (RelAbs 0.046 vs 0.051 on SCARED, 0.145 vs 0.153 on Hamlyn), so a different alignment convention could change or invert the comparison. Please state the exact alignment procedure and confirm that it was applied identically to every compared method.
  2. [Section IV-C, Eqs. (13)-(16)] The metric definitions appear to be mislabeled: Eq. (14) is the standard RMSE-log formula, Eq. (15) is the standard RMSE formula, and Eq. (13) appears to be a squared relative error averaged over pixels rather than the usual RelSq definition. The labels 'RMSE' and 'RMSElog' are also swapped relative to the variables used in the text. This ambiguity makes it difficult to reproduce the reported values and should be corrected.
  3. [Section IV-F, Table VI] The warm-up steps for DV-LoRA are evaluated on values 25000, 20000, 15000, and 5000, and the best value is selected from these results. No validation split is mentioned for this hyperparameter choice, and the tables do not report performance on a held-out validation set. Selecting the warm-up steps using the test set can inflate the reported test performance. Please describe how warm-up was chosen, or re-run the comparison with a validation-based selection.
  4. [Tables I, II, and V] All reported numbers appear to be from single runs, with no error bars, repeated trials, or significance tests. Several key comparisons are close: in Table I the gap to EndoDAC is 0.005 in RelAbs and 0.004 in delta, and several ablations in Table V differ by similar amounts. Without uncertainty quantification, the claim that the three-step training strategy is the source of the improvement is not statistically supported.
  5. [Section IV-D] The baseline reproduction protocol is described only as "some works with public codes or public weights are trained and tested based on the same setting." The paper does not report the exact training iterations, learning-rate schedules, loss weights, or finetuning details used for baselines such as AF-SfMLearner, IID-SfMLearner, MonoPCC, DVSMono, and the Depth Anything variants. Because the SOTA claim is a comparison against these methods, the reproducibility of the baseline setup is load-bearing and should be detailed, for example in an appendix or a supplementary table.
minor comments (7)
  1. [Section III-E] The heading contains a typo: "parametwr-efficient finetuning" should be "parameter-efficient finetuning."
  2. [Section IV-D] The sentence "Qualitative results in Fig further display the outstanding performance of the proposed method" has an incomplete figure reference and is then immediately repeated; it should refer to Fig. 4 and be stated once.
  3. [Tables I and II] The GC column is defined in the table caption as "Given Camera intrinsics," but Section IV-E also reports camera-intrinsics estimation. Please clarify whether the methods marked "Yes" use the ground-truth intrinsics during training, during evaluation, or both, and how this relates to the intrinsics prediction results.
  4. [Section III-D, Eq. (8)] The coordinate transformation in Eq. (8) is hard to parse: the notation C(·), the homogeneous coordinate representation, and the dimensions of the bracketed terms are not fully defined. A precise definition of each symbol would improve reproducibility.
  5. [Section III-E, Eq. (12)] In Eq. (12), the dimensions and roles of A, B, U, and V are not made explicit. The text says "ΛV and ΛU diagonal matrix of trainable weights U and V" while A and B are also trainable; please define all matrix shapes and state how the diagonal matrices are constructed.
  6. [Section IV-A] The description "all 92,672 endoscopic frames from the 21 rectified videos[28]" is ambiguous: is the count of frames from the subset used in [28]? Clarify the source of the rectified videos and the frame count.
  7. [Table V] The last row of Table V is garbled: "II→{I,III} Out of GPU Memory{I,II,III}" is not a clear description of the experiment. Please rephrase to state whether the run failed or was not performed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the SOTA claim rests on empirical comparisons and internal ablations, not on a derivation that reduces to its inputs.

full rationale

The paper's central claim is empirical: a three-step training strategy for self-supervised monocular depth achieves state-of-the-art results on SCARED and zero-shot results on Hamlyn. None of the reported metrics (Eqs. 13-16) are defined in terms of the model's training losses or parameters; they are external error measures against ground-truth depth. The training losses (Eqs. 5, 6, 11) supervise optical flow registration, intrinsic image decomposition, and rigid transformation alignment, respectively; they do not encode the final depth accuracy metric. The improvement of the multi-step strategy is supported by within-paper ablations (Table V) showing that the proposed step ordering and separation outperform fused variants, and by comparisons against methods from other groups (Yang et al., DVSMono, MonoPCC). The paper does cite prior work by overlapping authors: EndoDAC [16] shares co-author L. Bai and supplies the DV-LoRA adaptation module, and AF-SfMLearner [14] may be by the same first-author lineage and supplies appearance-flow components. However, these citations provide building blocks, not the load-bearing claim; the claimed benefit of the multi-step training is validated by ablation and by external baselines, not inferred from the citations themselves. The missing description of how scale-ambiguous monocular depth predictions are aligned to ground truth before computing RelAbs/RMSE/δ is a reproducibility gap, but it is not a circularity: the metric is not defined as a function of the model output by construction. No equation or training objective sets a predicted quantity equal to an input by definition. Therefore no circular step is identified.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on standard self-supervised photometric-consistency assumptions, the reliability of visibility masks and intrinsic image decomposition, and an unstated depth-scale alignment at evaluation time. The main hand-tuned parameters are the loss weights and the DV-LoRA warm-up (selected on the test set); no new physical entities are introduced.

free parameters (3)
  • Warm-up steps for DV-LoRA = 20000
    Chosen among 25000, 20000, 15000, 5000 by best SCARED test performance in Table VI; this is test-set selection.
  • Loss weighting coefficients = 0.001, 0.01, 0.02 (and ablation fusions 0.02, 0.02)
    Hand-chosen weights for smoothness, alignment and illumination-free losses (Eqs. 5, 11, 19, 20).
  • DV-LoRA rank = 4
    Set without reported ablation; affects finetuning capacity.
assumptions (4)
  • domain assumption Photometric consistency between target and warped source frames is a sufficient supervisory signal after adding appearance flow and intrinsic decomposition.
    Invoked in Eqs. 4, 6, 9, 10; assumes non-Lambertian tissue and lighting changes are captured by the extra modules.
  • domain assumption Visibility mask M from backward optical flow [26] correctly identifies occluded pixels.
    Used in Eqs. 4 and 9 to mask invalid warped pixels; if M is wrong, the losses train on invalid correspondences.
  • domain assumption The SCARED and Hamlyn preprocessing (320x256, split per [14]) and the evaluation protocol are fair and consistent across all compared methods.
    Section IV-A and IV-C define datasets and metrics, but depth-scale alignment for monocular predictions is not described.
  • domain assumption A foundation model pretrained on natural images provides a useful depth prior for endoscopy.
    Section III-E assumes the encoder of Depth Anything is a good starting point and only LoRA adapters must be trained.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EndoMUST: Monocular Depth Estimation for Robotic Endoscopy via End-to-end Multi-step Self-supervised Training." pith.science (2026). https://pith.science/paper/BOV4Y3CW

@misc{pith2026250616017,
  author       = {Pith},
  title        = {Pith review of: EndoMUST: Monocular Depth Estimation for Robotic Endoscopy via End-to-end Multi-step Self-supervised Training},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BOV4Y3CW}},
  note         = {Machine review of arXiv:2506.16017}
}
abstract

Monocular depth estimation and ego-motion estimation are significant tasks for scene perception and navigation in stable, accurate and efficient robot-assisted endoscopy. To tackle lighting variations and sparse textures in endoscopic scenes, multiple techniques including optical flow, appearance flow and intrinsic image decomposition have been introduced into the existing methods. However, the effective training strategy for multiple modules are still critical to deal with both illumination issues and information interference for self-supervised depth estimation in endoscopy. Therefore, a novel framework with multistep efficient finetuning is proposed in this work. In each epoch of end-to-end training, the process is divided into three steps, including optical flow registration, multiscale image decomposition and multiple transformation alignments. At each step, only the related networks are trained without interference of irrelevant information. Based on parameter-efficient finetuning on the foundation model, the proposed method achieves state-of-the-art performance on self-supervised depth estimation on SCARED dataset and zero-shot depth estimation on Hamlyn dataset, with 4\%$\sim$10\% lower error. The evaluation code of this work has been published on https://github.com/BaymaxShao/EndoMUST.

Figures

Figures reproduced from arXiv: 2506.16017 by the authors.

Figure 1
Figure 1. Challenges for depth estimation in robotic endoscopy. Highlights [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The overview of the proposed method. Each end-to-end training epoch is divided into three steps. In Step I, the optical flow prediction network [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Parameter-efficient finetuning in each Transformer block based on [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Qualitative Results of Depth Estimation. Results of self-supervised depth estimation on SCARED dataset are in first three rows. Results of [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Qualitative results of ego-motion estimation. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Qualitative Comparison of Ablation Studies on SCARED dataset. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 23 canonical work pages

  1. [1]

    Real-time navigation for laparoscopic hepatectomy using image fusion of preoperative 3d surgical plan and intraoperative indocyanine green fluorescence imaging,

    P. Zhang, H. Luo, W. Zhu, J. Yang, N. Zeng, Y . Fan, S. Wen, N. Xiang, F. Jia, and C. Fang, “Real-time navigation for laparoscopic hepatectomy using image fusion of preoperative 3d surgical plan and intraoperative indocyanine green fluorescence imaging,” Surgical endoscopy, vol. 34, pp. 3449–3459, 2020

    work page 2020

  2. [2]

    Self-supervised lightweight depth estimation in endoscopy combining cnn and trans- former,

    Z. Yang, J. Pan, J. Dai, Z. Sun, and Y . Xiao, “Self-supervised lightweight depth estimation in endoscopy combining cnn and trans- former,” IEEE Transactions on Medical Imaging , vol. 43, no. 5, pp. 1934–1944, 2024

    work page 1934

  3. [3]

    Bdis: Bayesian dense inverse searching method for real-time stereo surgical image matching,

    J. Song, Q. Zhu, J. Lin, and M. Ghaffari, “Bdis: Bayesian dense inverse searching method for real-time stereo surgical image matching,” IEEE Transactions on Robotics , vol. 39, no. 2, pp. 1388–1406, 2022

    work page 2022

  4. [4]

    Msdesis: Multitask stereo disparity estimation and surgical instrument segmentation,

    D. Psychogyios, E. Mazomenos, F. Vasconcelos, and D. Stoyanov, “Msdesis: Multitask stereo disparity estimation and surgical instrument segmentation,” IEEE transactions on medical imaging, vol. 41, no. 11, pp. 3218–3230, 2022

    work page 2022

  5. [5]

    Laparoscopic stereo matching using 3-dimensional fourier transform with full multi- scale features,

    R. Wu, P. Liang, Y . Liu, Y . Huang, W. Li, and Q. Chang, “Laparoscopic stereo matching using 3-dimensional fourier transform with full multi- scale features,” Engineering Applications of Artificial Intelligence, vol. 139, p. 109654, 2025

    work page 2025

  6. [6]

    Cgi-stereo: Accurate and real-time stereo matching via context and geometry interaction,

    G. Xu, H. Zhou, and X. Yang, “Cgi-stereo: Accurate and real-time stereo matching via context and geometry interaction,” arXiv preprint arXiv:2301.02789, 2023

    arXiv 2023

  7. [7]

    Dual cnn models for unsupervised monocular depth estimation,

    V . K. Repala and S. R. Dubey, “Dual cnn models for unsupervised monocular depth estimation,” in Pattern Recognition And Machine Intelligence: 8th International Conference, PReMI 2019, Tezpur, India, December 17-20, 2019, Proceedings, Part I. Springer, 2019, pp. 209– 217

    work page 2019

  8. [8]

    Structured attention guided convolutional neural fields for monocular depth estimation,

    D. Xu, W. Wang, H. Tang, H. Liu, N. Sebe, and E. Ricci, “Structured attention guided convolutional neural fields for monocular depth estimation,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2018, pp. 3917–3925

    work page 2018

Show all 30 references
  1. [9]

    Unsupervised reverse domain adaptation for synthetic medical images via adversarial training,

    F. Mahmood, R. Chen, and N. J. Durr, “Unsupervised reverse domain adaptation for synthetic medical images via adversarial training,” IEEE transactions on medical imaging, vol. 37, no. 12, pp. 2572–2581, 2018

    2018

  2. [10]

    Slam endoscopy enhanced by adversarial depth prediction,

    R. J. Chen, T. L. Bobrow, T. Athey, F. Mahmood, and N. J. Durr, “Slam endoscopy enhanced by adversarial depth prediction,” arXiv preprint arXiv:1907.00283, 2019

    1907 arXiv

  3. [11]

    Monovit: Self-supervised monocular depth estimation with a vision transformer,

    C. Zhao, Y . Zhang, M. Poggi, F. Tosi, X. Guo, Z. Zhu, G. Huang, Y . Tang, and S. Mattoccia, “Monovit: Self-supervised monocular depth estimation with a vision transformer,” in 2022 international conference on 3D vision (3DV) . IEEE, 2022, pp. 668–678

    2022

  4. [12]

    Lite-mono: A lightweight cnn and transformer architecture for self-supervised monocular depth estimation,

    N. Zhang, F. Nex, G. V osselman, and N. Kerle, “Lite-mono: A lightweight cnn and transformer architecture for self-supervised monocular depth estimation,” in Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition , 2023, pp. 18 537– 18 546

    2023

  5. [13]

    Monod- iffusion: self-supervised monocular depth estimation using diffusion model,

    S. Shao, Z. Pei, W. Chen, D. Sun, P. C. Chen, and Z. Li, “Monod- iffusion: self-supervised monocular depth estimation using diffusion model,” IEEE Transactions on Circuits and Systems for Video Tech- nology, 2024

    2024

  6. [14]

    Self- supervised monocular depth and ego-motion estimation in endoscopy: Appearance flow to the rescue,

    S. Shao, Z. Pei, W. Chen, W. Zhu, X. Wu, D. Sun, and B. Zhang, “Self- supervised monocular depth and ego-motion estimation in endoscopy: Appearance flow to the rescue,” Medical image analysis , vol. 77, p. 102338, 2022

    2022

  7. [15]

    Image intrinsic-based unsupervised monocular depth estimation in endoscopy,

    B. Li, B. Liu, M. Zhu, X. Luo, and F. Zhou, “Image intrinsic-based unsupervised monocular depth estimation in endoscopy,”IEEE Journal of Biomedical and Health Informatics , pp. 1–11, 2024

    2024

  8. [16]

    Endodac: Efficient adapting foundation model for self-supervised depth estimation from any endoscopic camera,

    B. Cui, M. Islam, L. Bai, A. Wang, and H. Ren, “Endodac: Efficient adapting foundation model for self-supervised depth estimation from any endoscopic camera,” in International Conference on Medical Image Computing and Computer-Assisted Intervention . Springer, 2024, pp. 208–218

    2024

  9. [17]

    Unsupervised odometry and depth learn- ing for endoscopic capsule robots,

    M. Turan, E. P. Ornek, N. Ibrahimli, C. Giracoglu, Y . Almalioglu, M. F. Yanik, and M. Sitti, “Unsupervised odometry and depth learn- ing for endoscopic capsule robots,” in 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2018, pp. 1801– 1807

    2018

  10. [18]

    Dense depth estimation in monocular endoscopy with self-supervised learning methods,

    X. Liu, A. Sinha, M. Ishii, G. D. Hager, A. Reiter, R. H. Taylor, and M. Unberath, “Dense depth estimation in monocular endoscopy with self-supervised learning methods,” IEEE transactions on medical imaging, vol. 39, no. 5, pp. 1438–1447, 2019

    2019

  11. [19]

    Unsupervised-learning-based continuous depth and motion estimation with monocular endoscopy for virtual reality minimally invasive surgery,

    L. Li, X. Li, S. Yang, S. Ding, A. Jolfaei, and X. Zheng, “Unsupervised-learning-based continuous depth and motion estimation with monocular endoscopy for virtual reality minimally invasive surgery,” IEEE transactions on industrial informatics , vol. 17, no. 6, pp. 3920–3928, 2020

    2020

  12. [20]

    Endoslam dataset and an unsupervised monocular visual odometry and depth estimation approach for endoscopic videos,

    K. B. Ozyoruk, G. I. Gokceler, T. L. Bobrow, G. Coskun, K. Incetan, Y . Almalioglu, F. Mahmood, E. Curto, L. Perdigoto, M. Oliveiraet al., “Endoslam dataset and an unsupervised monocular visual odometry and depth estimation approach for endoscopic videos,” Medical image analys...

    2021

  13. [21]

    Self-supervised monocular depth estimation for gastrointestinal endoscopy,

    Y . Liu and S. Zuo, “Self-supervised monocular depth estimation for gastrointestinal endoscopy,” Computer Methods and Programs in Biomedicine, vol. 238, p. 107619, 2023

    2023

  14. [22]

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

    L. Yang, B. Kang, Z. Huang, X. Xu, J. Feng, and H. Zhao, “Depth anything: Unleashing the power of large-scale unlabeled data,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 10 371–10 381

    2024

  15. [23]

    Depth anything v2,

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

    2024 arXiv

  16. [24]

    Improved self-supervised monocular endoscopic depth estimation based on pose alignment-friendly dynamic view selection,

    Y . Zhou, S. He, H. Wang, F. Huang, M. Liu, Q. Li, and Z. Wang, “Improved self-supervised monocular endoscopic depth estimation based on pose alignment-friendly dynamic view selection,” in 2024 IEEE International Conference on Bioinformatics and Biomedicine (BIBM). IEEE, 2024,...

    2024

  17. [25]

    Spatial transformer networks,

    M. Jaderberg, K. Simonyan, A. Zisserman et al., “Spatial transformer networks,” Advances in neural information processing systems, vol. 28, 2015

    2015

  18. [26]

    Occlusion aware unsupervised learning of optical flow,

    Y . Wang, Y . Yang, Z. Yang, L. Zhao, P. Wang, and W. Xu, “Occlusion aware unsupervised learning of optical flow,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2018, pp. 4884–4893

    2018

  19. [27]

    Monopcc: Photometric-invariant cycle constraint for monocular depth estimation of endoscopic images,

    Z. Wang, Y . Zhou, S. He, T. Li, F. Huang, Q. Ding, X. Feng, M. Liu, and Q. Li, “Monopcc: Photometric-invariant cycle constraint for monocular depth estimation of endoscopic images,” arXiv preprint arXiv:2404.16571, 2024

    2024 arXiv

  20. [28]

    Endo-depth-and-motion: Reconstruction and tracking in endoscopic videos using depth networks and photometric constraints,

    D. Recasens, J. Lamarca, J. M. F ´acil, J. M. Montiel, and J. Civera, “Endo-depth-and-motion: Reconstruction and tracking in endoscopic videos using depth networks and photometric constraints,” IEEE Robotics and Automation Letters , vol. 6, no. 4, pp. 7225–7232, 2021

    2021

  21. [29]

    Stereo correspondence and reconstruction of endoscopic data challenge,

    M. Allan, J. Mcleod, C. Wang, J. C. Rosenthal, Z. Hu, N. Gard, P. Eisert, K. X. Fu, T. Zeffiro, W. Xia et al., “Stereo correspondence and reconstruction of endoscopic data challenge,” arXiv preprint arXiv:2101.01133, 2021

    2021 arXiv

  22. [30]

    Hamlyn centre laparoscopic / endoscopic video datasets

    S. Giannarou, D. Stoyanov, D. Noonan, G. Mylonas, J. Clark, M. Visentini-Scarzanella, P. Mountney, and G.-Z. Yang, “Hamlyn centre laparoscopic / endoscopic video datasets.” [Online]. Available: https://hamlyn.doc.ic.ac.uk/vision/

Pith tools

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