Pith. sign in

REVIEW 4 major objections 5 minor 46 references

EndoMD-SLAM: Endoscopic Gaussian Splatting SLAM under Optical Degradation with Memory and Static-Transient Decomposition

T0 review · 4 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read EndoMD-SLAM keeps colonoscopy mapping stable under lens contamination by gating map updates and separating transient artifacts from the anatomical field.

desk verdict EndoMD-SLAM tackles a real, under-addressed failure mode in endoscopic 3DGS SLAM with a sensible gated-memory plus static-transient decomposition design, but the evaluation is too thin to fully trust the headline numbers. read the letter →

arxiv 2608.08949 v1 pith:ZR4FFTHL submitted 2026-08-09 cs.CV

Nuo Chen , Kangqi Ni , Lulin Liu , Joga Ivatury , Ying Ding , Farshid Alambeigi , Tianlong Chen , Zhiwen Fan This is my paper · ORCID
classification cs.CV
keywords endoscopicSLAMGaussiansplattingopticaldegradationstatic-transientdecompositiontemporalmemoryrelocalizationcolonoscopy3DreconstructionRGB-D
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

EndoMD-SLAM is an online dense-mapping system for colonoscopy built on Gaussian splatting, and this paper claims it keeps tracking and 3D reconstruction stable when the lens is intermittently fouled by water droplets, debris, or flushing. Standard splatting SLAM assumes clean multi-view photometric consistency, so it fuses camera-attached contaminants into the persistent map and lets corrupted poses accumulate. The proposed system instead checks each frame with a reliability gate, suspends map updates when the gate closes, recovers the pose by matching against a memory bank of clean keyframes, and factorizes the scene into a persistent static field and a per-frame transient field that absorbs the artifacts. On ten severely degraded colonoscopy sequences the authors report an absolute trajectory error of 3.02 mm and a rendering PSNR of 18.06 dB; compared with the endoscopy-specific 3DGS baseline, that is roughly a 91% reduction in trajectory error and a 9.9 dB gain in PSNR. The contribution, read in good faith, is a demonstration that degradation robustness in online endoscopic SLAM is achievable by making the decision to trust a frame explicit and by keeping transient occluders out of the anatomical map.

What carries the argument

The load-bearing object is a paired trust-and-factorization mechanism: the reliability gate and the static-transient Gaussian fields. The gate is $g_t=\mathbb{I}(\rho_t\ge\tau_\rho \wedge N_{\mathrm{valid}}\ge\tau_n)$, where $\rho_t$ and $N_{\mathrm{valid}}$ come from a weighted Procrustes pose solve on foundation-model correspondences, with weights $w'_i=w_i\max(1-\hat{\alpha}^t_{t-1}(u_i),0)^\gamma$ modulated by the rendered transient opacity. When $g_t=1$ the static map is updated; when $g_t=0$ map updates are suspended, degraded frames are allowed transient-only fitting, and the pose is recovered by selecting the memory-bank keyframe with the highest geometric consensus $\rho_t^{(k)}$. The representation is two Gaussian fields composited as $\hat{I}_t=\hat{I}^t_t+(1-\hat{\alpha}^t_t)\odot\hat{I}^s_t$, with regularization pushing transient occupancy up and static emptiness down. Together these keep corrupted poses and camera-attached artifacts from entering the persistent anatomy.

What would settle it

Render synthetic water droplets and debris with known ground-truth masks onto clean colonoscopy frames with known poses, run EndoMD-SLAM with and without the transient-alpha weighting, and compare the predicted transient field against the known masks. If the predicted mask has low overlap with the true contamination yet the trajectory error stays near 3 mm, or if removing the weighting leaves the pose error nearly unchanged, then the decomposition is not the active ingredient and the memory gate is carrying the result.

Watch

Extended reading notes

Core claim

The paper's central discovery is that the two failure modes induced by optical degradation—tracking drift during flushing and irreversible fusion of droplets into geometry—are both consequences of the SLAM system treating every observation as reliable new anatomy. The fix is to make trust explicit and to split the representation. A binary reliability gate, computed from the fraction of geometrically valid correspondences after weighting by the transient opacity mask, decides whether a frame may write into the static map; when the gate closes, the system suspends static updates and relocalizes using candidate keyframes retrieved from a temporal memory bank. In parallel, the scene is represented as two Gaussian fields—a global static field for anatomy and a per-frame transient field for contaminants—composited by occlusion-style alpha blending, so the optimization can reproduce the degraded view while the static field stays clean. The paper's reported outcome of this design is an ATE of 3.02 mm and PSNR of 18.06 dB on the degraded benchmark, against 35.91 mm and 8.16 dB for the endoscopy-specific baseline it aims to replace.

Load-bearing premise

The whole tracking chain assumes that, after the transient mask has down-weighted contaminated regions, the learned matcher still finds enough correct 3D point pairs on smooth, texture-poor colon walls to recover the pose; the paper reports no direct measurement of mask accuracy or surviving correspondence density on clean tissue.

Editorial extensions

If this is right

  • If the central claim holds, a dense colonoscopy SLAM system can survive repeated lens flushing and droplet occlusion without accumulating pose error, which is the precondition for using live 3D reconstruction as a navigation aid.
  • Frames deemed unreliable are never written into the static map, so degraded intervals cannot permanently corrupt the anatomy; the final reconstruction remains usable for documentation even when the source video is badly contaminated.
  • Because the scene is factored into static and transient fields, the same model can render both a clean anatomical view and the true degraded view, giving the clinician both representations at once.
  • The transient field is recycled per-frame, so the cost of artifact handling does not grow with the number of contamination events, keeping mapping time nearly flat over a long procedure.

Reading between the lines

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

  • The paper evaluates on the C3VDv2 benchmark only and states that strictly real-time clinical deployment remains future work; at 0.95 FPS the current system is not yet live-grade, and generalizing to real colonoscopy video is untested.
  • A diagnostic the authors do not run is to measure the transient field against ground-truth contamination masks; the method's logic predicts that mask accuracy should track pose accuracy, and that correlation would isolate which component produces the gain.
  • The same gating-plus-decomposition pattern should transfer to other camera-attached distractors in surgery, such as smoke, blood, or tool occlusion, because the transient field is agnostic to what it absorbs; this is an editorial extension.
  • The reported results assume RGB-D input, and the transient field is initialized by backprojecting depth, so in a monocular deployment the accuracy of the depth prior would likely bound how cleanly small contaminants are separated.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes EndoMD-SLAM, a 3D Gaussian Splatting SLAM system for colonoscopy under optical degradation. The method combines a temporal memory bank that gates static map updates and triggers relocalization when tracking is unreliable, with a static-transient Gaussian decomposition that separates persistent anatomy from camera-attached contaminants. The paper curates a ten-sequence benchmark from C3VDv2 with 'water on lens' and 'debris on lens' metadata, and reports that EndoMD-SLAM achieves an ATE of 3.02 mm and PSNR of 18.06 dB, outperforming NICE-SLAM, MonoGS, and EndoGSLAM, with an ablation study attributing the gains to both the memory and the decomposition modules.

Significance. The problem addressed is important and clinically relevant: standard dense SLAM systems do fail when lens contamination and water flushing violate photometric consistency, and the proposed memory-gating plus static-transient decomposition is a sensible and well-motivated design. The curated degradation-focused benchmark is a useful resource, and the qualitative results suggest the method can prevent map corruption on the tested sequences. However, the experimental evidence as presented is not yet sufficient to support the central robustness claims: the ablation baseline is numerically identical to a competitor system, there are no error bars or per-sequence statistics, hyperparameters appear to be selected on the same ten sequences that are then used for evaluation, and the tracking pipeline's reliance on MASt3R correspondences plus a learned transient mask is not directly validated. With additional experiments and analysis these concerns are addressable; the core idea is defensible.

major comments (4)
  1. [Section IV-C, Table II] The row labelled 'Baseline' in Table II reports exactly the same numbers as EndoGSLAM in Table I (PSNR 8.16, SSIM 0.335, LPIPS 0.628, RMSE 26.21, ATE 35.91). Since EndoMD-SLAM is described as using a MASt3R-based foundation-model tracker in Section III-A-a, this row cannot be the proposed architecture without the memory and decomposition modules unless the authors are equating EndoMD-SLAM's baseline with EndoGSLAM. As written, the ablation does not isolate the contribution of the proposed components on the proposed architecture, and the conclusion that both components are 'critical' is not supported. Please re-run the ablation from the full EndoMD-SLAM pipeline with each component removed and report the no-component baseline using the same tracker and mapping backbone.
  2. [Section IV-A-d, Table I] All reported metrics are aggregate numbers over ten sequences, with no per-sequence breakdown, standard deviation, or failure-case analysis. Moreover, the hyperparameters in Section IV-A-d (tau_rho=0.15, tau_n=500, gamma=2.0, N_t=1000, |C_t|=10, lambda_rgb=1.0, lambda_d=1.0, w_tr=0.01, w_bg=0.01) are listed without any validation procedure. If these thresholds and loss weights were tuned on the same ten sequences used in Table I, the headline margins (91% ATE reduction, 9.9 dB PSNR improvement) may partly reflect tuning on the test set. Please report per-sequence metrics, error bars, and either a held-out validation split or a sensitivity analysis over the key thresholds.
  3. [Section III-A-a, Eqs. (1)-(3)] The tracking chain rests entirely on MASt3R correspondences being re-weighted by the transient mask in Eq. (1), followed by weighted Procrustes in Eq. (2) and the reliability gate in Eq. (3). The paper provides no quantitative evidence that (a) MASt3R supplies enough metric 3D correspondences in textureless colon tissue, or (b) the transient mask suppresses the high-confidence matches that MASt3R assigns to moving lens contaminants. Without such evidence, the thresholds tau_rho=0.15 and tau_n=500 are unjustified and the robustness claim may not generalize beyond the ten curated sequences. Please report correspondence density, inlier fractions, mask-quality metrics (e.g., overlap of transient opacity with contaminant regions), and the distribution of rho and N_valid on clean versus degraded frames.
  4. [Section III-A-a and III-B] There is a circular dependency between tracking and decomposition: the transient mask used to re-weight correspondences in Eq. (1) is learned from poses produced by the same tracking loop, and the static map is updated only when the pose is deemed reliable. The paper does not explain how the transient field is initialized or maintained on severely degraded frames when the pose is gated out, nor does it analyze whether a wrong pose corrupts the mask and vice versa. Please describe the initialization procedure for degraded frames and provide an experiment that varies the transient-field initialization or uses a warm-start, so that the stability of the coupled loop can be assessed.
minor comments (5)
  1. [Abstract, Section IV-B] The 91% ATE reduction and 9.9 dB PSNR improvement are relative to EndoGSLAM, which has the worst ATE among the compared baselines; please state the comparison baseline explicitly in the abstract to avoid over-generalization.
  2. [Section IV-A-d] The method uses RGB-D input, but the abstract and introduction describe the contribution as supporting clinical endoscopic navigation without noting this requirement; please state clearly that depth input is needed and discuss how this maps to available colonoscopy hardware.
  3. [Section III-B, Eq. (7)-(8)] The depth loss supervises only static geometry, but during lens contamination the measured depth itself may be corrupted; please clarify how the depth supervision handles degraded frames where the depth observation does not correspond to anatomy.
  4. [Section IV-D] There is a typo in Section IV-D: 'estimatsed' should be 'estimated'.
  5. [Figure 3] It would strengthen the paper to report a quantitative measure of transient-field separation, such as the fraction of transient opacity overlapping the contaminant regions, instead of relying only on qualitative visualizations.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the headline gains are empirical benchmark comparisons, not quantities derived from fitted inputs or self-citations.

full rationale

EndoMD-SLAM is an empirical SLAM system; there is no formal derivation chain whose conclusions are fixed by its assumptions. The headline ATE reduction (35.91 mm to 3.02 mm) and PSNR gain (8.16 to 18.06 dB) are measured against external baselines on the C3VDv2-derived benchmark, and the ablation attributes the gains to the proposed gating and decomposition. I find no step in which a reported quantity equals a fitted input or a cited result by construction. The tracking loop (Eqs. 1-2) uses MASt3R correspondences reweighted by the transient mask, and the transient field is initialized from depth backprojection (Eq. 9); this is a coupled bootstrapping optimization, not a tautology, and the paper explicitly defaults the mask to 1 before transient initialization. The reliability gate thresholds and loss weights are stated as implementation details, but the paper does not show they were selected by fitting to the reported metrics, so the evaluation concern is potential overfitting, not definitional circularity. One genuine evaluation weakness is that the 'Baseline' row in Table II is numerically identical to the EndoGSLAM row in Table I (8.16 PSNR, 26.21 RMSE, 35.91 ATE), even though EndoMD-SLAM uses a different MASt3R-based tracker; this undermines the component-level attribution but does not make the main benchmark comparison circular, because the full system is still independently evaluated against published baselines. I therefore find no significant circularity, and score 0.

Assumptions & free parameters 11 free parameters · 4 assumptions · 1 invented entities

The central claim rests on a small set of hand-tuned hyperparameters (thresholds and loss weights), on the accuracy of the C3VDv2 ground truth and tags, on the reliability of MASt3R in textureless endoscopy, and on the assumption that a per-frame transient field cleanly separates camera-attached artifacts from static anatomy. No external independent evidence is provided for the transient field representation.

free parameters (11)
  • reliability gate match fraction threshold tau_rho = 0.15
    Used in Eq. (3) to decide whether a frame is reliable; value given in Section IV-A-d without sensitivity analysis.
  • reliability gate valid point threshold tau_n = 500
    Minimum number of valid metric correspondences in Eq. (3); chosen in Section IV-A-d without validation split.
  • transient weight modulation exponent gamma = 2.0
    Exponent in Eq. (1) that shapes the transient-mask weighting; set in Section IV-A-d.
  • transient Gaussians per frame N_t = 1000
    Fixed capacity of the per-frame transient field in Section III-B-c and IV-A-d.
  • reconstruction loss weight lambda_rgb = 1.0
    Weight in Eq. (7); given in Section IV-A-d.
  • depth loss weight lambda_d = 1.0
    Weight in Eq. (7); given in Section IV-A-d.
  • transient occupancy regularizer weight w_tr = 0.01
    Weight for the transient alpha penalty in Eq. (8); given in Section IV-A-d.
  • static coverage regularizer weight w_bg = 0.01
    Weight for the static background penalty in Eq. (8); given in Section IV-A-d.
  • number of retrieval candidates |C_t| = 10
    Number of historical keyframes queried during relocalization; set in Section IV-A-d.
  • transient recycling threshold tau_act = not specified
    Threshold in Eq. (10) that determines when a transient Gaussian is inactive; defined but never assigned a numeric value.
  • mapping iterations per frame = 25
    Number of optimization iterations per frame for static and transient fields; set in Section IV-A-d.
assumptions (4)
  • domain assumption C3VDv2 provides accurate depth and pose ground truth, and its metadata tags correctly identify degraded sequences.
    Every metric in Tables I-III depends on the correctness of this dataset and its degradation tags; the paper provides no independent validation of the tags.
  • domain assumption MASt3R foundation model provides reliable 3D correspondences in textureless endoscopic images, and the transient mask can suppress contamination matches.
    Tracking in Eq. (2) relies entirely on these correspondences; Section III-A-a states the model 'operates under a general static-scene assumption' and may match contamination.
  • domain assumption Camera-attached artifacts are well represented by a per-frame transient Gaussian field that does not entangle with static anatomy.
    The decomposition in Section III-B assumes that moving droplets and debris can be captured by a fixed-capacity per-frame field with recycling; the paper does not test scenarios where artifacts persist across many frames.
  • domain assumption The colon tissue is static during mapping, aside from the transient artifacts.
    The static map in Eq. (5) is a single global field; deformable tissue or tool motion would violate this assumption and is outside the paper's scope.
invented entities (1)
  • Per-frame transient Gaussian field
    purpose: Absorbs lens contaminants such as water droplets and debris so they are not fused into the persistent static anatomy.
    This is a modeling construct introduced by the paper. Its effectiveness is measured only through rendering quality on the same benchmark; there is no external falsifiable handle, such as a predicted physical quantity that could be confirmed elsewhere.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EndoMD-SLAM: Endoscopic Gaussian Splatting SLAM under Optical Degradation with Memory and Static-Transient Decomposition." pith.science (2026). https://pith.science/paper/ZR4FFTHL

@misc{pith2026260808949,
  author       = {Pith},
  title        = {Pith review of: EndoMD-SLAM: Endoscopic Gaussian Splatting SLAM under Optical Degradation with Memory and Static-Transient Decomposition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZR4FFTHL}},
  note         = {Machine review of arXiv:2608.08949}
}
read the original abstract

Dense 3D reconstruction is critical for clinical endoscopic navigation and documentation. While Gaussian Splatting SLAM systems show promise in this domain, they fundamentally rely on strict multi-view photometric consistency. In routine procedures, this assumption is severely violated by intermittent optical degradations like moving debris and water flushing. Standard systems erroneously fuse these cameraattached artifacts into the persistent 3D geometry, causing severe tracking drift and irreversible map corruption. To address this limitation, we propose EndoMD-SLAM, a framework designed to maintain stability under optical degradation through specialized tracking and mapping mechanisms. On the tracking side, a memory-driven gating mechanism detects unreliable observations to suspend map updates and utilizes historical keyframes for drift-aware relocalization. On the mapping side, a self-supervised static-transient decomposition isolates visual contaminants into a dedicated transient field. This explicit separation prevents artifacts from structurally entangling with the persistent anatomical map. We curate a degradationfocused benchmark from colonoscopy videos to systematically evaluate these failure modes. Extensive experiments show that while standard baselines fail under severe optical degradation, EndoMD-SLAM preserves geometric integrity, reducing absolute trajectory error by 91% and improving rendering fidelity by 9.9 dB PSNR.

Figures

Figures reproduced from arXiv: 2608.08949 by the authors.

Figure 1
Figure 1. SLAM under severe optical degradation. Standard Gaussian Splatting SLAM systems fail under routine optical degradations, suffering severe tracking drift and erroneously mapping updates. With the proposed memory and decompo￾sition mechanisms, EndoMD-SLAM isolates these degrada￾tions caused by contaminants and water flushing, maintaining robust online tracking and recovering a clean, high-fidelity 3D anatomical recons… view at source ↗
Figure 2
Figure 2. Section III-A introduces the temporal memory for [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 2
Figure 2. Overview of the EndoMD-SLAM architecture. Input frames are first evaluated by a reliability gate. Non-degraded observations proceed to tracking optimization and update the global static map. Degraded frames activate the temporal memory, which queries a keyframe memory bank to perform drift-aware relocalization while safely suspending static map updates. Concurrently, a static-transient decomposition explicitly facto… view at source ↗
Figures from the paper (4 more)
Figure 3
Figure 3. Figure 3: Qualitative comparison of novel view synthesis and field decomposition under severe optical degradation. Standard baselines (EndoGSLAM, MonoGS, NICE-SLAM) exhibit prominent rendering artifacts and geometric distortions under degraded visual evidence. EndoMD-SLAM explic…
Figure 4
Figure 4. Figure 4: Qualitative comparison of estimated camera tra￾jectories across diverse colonoscopy sequences. Under severe lens contamination, the baseline method (blue dashed lines) experiences catastrophic tracking drift and frequent localization failures due to the sudden loss of …
Figure 5
Figure 5. Figure 5: Under severe optical degradation, EndoGSLAM ex [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 5
Figure 5. Figure 5: Qualitative 3D reconstruction of global colon anatomy under severe lens contamination. The top row (EndoGSLAM) exhibits catastrophic structural breakdown and geometric distortion caused by the naive fusion of transient artifacts and tracking drift. Conversely, EndoMD￾S…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 28 canonical work pages

  1. [1]

    Ad- vances in real-time 3d reconstruction for medical endoscopy,

    A. Richter, T. Steinmann, J.-C. Rosenthal, and S. J. Rupitsch, “Ad- vances in real-time 3d reconstruction for medical endoscopy,”Journal of imaging, vol. 10, no. 5, p. 120, 2024

    work page 2024

  2. [2]

    Frontiers of robotic colonoscopy: A comprehensive review of robotic colonoscopes and technologies,

    G. Ciuti, K. Skonieczna- ˙Zydecka, W. Marlicz, V . Iacovacci, H. Liu, D. Stoyanov, A. Arezzo, M. Chiurazzi, E. Toth, H. Thorlacius,et al., “Frontiers of robotic colonoscopy: A comprehensive review of robotic colonoscopes and technologies,”Journal of clinical medicine, vol. 9, no. 6, p. 1648, 2020

    work page 2020

  3. [3]

    Endomapper dataset of complete calibrated endoscopy proce- dures,

    P. Azagra, C. Sostres, ´A. Ferr ´andez, L. Riazuelo, C. Tomasini, O. L. Barbed, J. Morlana, D. Recasens, V . M. Batlle, J. J. G´omez-Rodr´ıguez, et al., “Endomapper dataset of complete calibrated endoscopy proce- dures,”Scientific Data, vol. 10, no. 1, p. 671, 2023

    work page 2023

  4. [4]

    Endogslam: Real-time dense reconstruction and tracking in endoscopic surgeries using gaussian splatting,

    K. Wang, C. Yang, Y . Wang, S. Li, Y . Wang, Q. Dou, X. Yang, and W. Shen, “Endogslam: Real-time dense reconstruction and tracking in endoscopic surgeries using gaussian splatting,” inInternational Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2024, pp. 219–229

    work page 2024

  5. [5]

    Endoflow-slam: Real-time endoscopic slam with flow-constrained gaussian splatting,

    T. Wu, Y . Miao, Z. Li, H. Zhao, K. Dang, J. Su, L. Yu, and H. Li, “Endoflow-slam: Real-time endoscopic slam with flow-constrained gaussian splatting,” inInternational Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2025, pp. 202–212

    work page 2025

  6. [6]

    Sage: slam with appearance and geometry prior for endoscopy,

    X. Liu, Z. Li, M. Ishii, G. D. Hager, R. H. Taylor, and M. Unberath, “Sage: slam with appearance and geometry prior for endoscopy,” in 2022 International conference on robotics and automation (ICRA). IEEE, 2022, pp. 5587–5593

    work page 2022

  7. [7]

    Coarse-to-fine real-time gs-slam from endoscopic rgbd data,

    Z. Cao, Y . Zhang, Y . Qian, and Q. Wang, “Coarse-to-fine real-time gs-slam from endoscopic rgbd data,” inSeventeenth International Conference on Digital Image Processing (ICDIP 2025), vol. 13709. SPIE, 2025, pp. 210–219

    work page 2025

  8. [8]

    Pr-endo: Physically based relightable gaus- sian splatting for endoscopy,

    J. Kaleta, W. Smolak-Dy ˙zewska, D. Malarz, D. Dall’Alba, P. Ko- rzeniowski, and P. Spurek, “Pr-endo: Physically based relightable gaus- sian splatting for endoscopy,” inInternational Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2025, pp. 391–401

    work page 2025

Show all 46 references
  1. [9]

    Visual slam for handheld monocular endoscope,

    O. G. Grasa, E. Bernal, S. Casado, I. Gil, and J. Montiel, “Visual slam for handheld monocular endoscope,”IEEE transactions on medical imaging, vol. 33, no. 1, pp. 135–146, 2013

    2013

  2. [10]

    Slam based quasi dense reconstruction for min- imally invasive surgery scenes,

    N. Mahmoud, A. Hostettler, T. Collins, L. Soler, C. Doignon, and J. M. M. Montiel, “Slam based quasi dense reconstruction for min- imally invasive surgery scenes,”arXiv preprint arXiv:1705.09107, 2017

    2017 arXiv

  3. [11]

    Visual slam for bronchoscope tracking and bronchus reconstruction in bronchoscopic navigation,

    C. Wang, M. Oda, Y . Hayashi, T. Kitasaka, H. Honma, H. Takabatake, M. Mori, H. Natori, and K. Mori, “Visual slam for bronchoscope tracking and bronchus reconstruction in bronchoscopic navigation,” in Medical Imaging 2019: Image-Guided Procedures, Robotic Interven- tions, and ...

    2019

  4. [12]

    Sd-defslam: Semi-direct monocular slam for deformable and intracorporeal scenes,

    J. J. G ´omez-Rodr´ıguez, J. Lamarca, J. Morlana, J. D. Tard´os, and J. M. Montiel, “Sd-defslam: Semi-direct monocular slam for deformable and intracorporeal scenes,” in2021 IEEE international conference on robotics and automation (ICRA). IEEE, 2021, pp. 5170–5177

    2021

  5. [13]

    Oneslam to map them all: a generalized approach to slam for monocular endoscopic imaging based on tracking any point,

    T. Teufel, H. Shu, R. D. Soberanis-Mukul, J. E. Mangulabnan, M. Sahu, S. S. Vedula, M. Ishii, G. Hager, R. H. Taylor, and M. Unberath, “Oneslam to map them all: a generalized approach to slam for monocular endoscopic imaging based on tracking any point,” International Journal ...

    2024

  6. [14]

    Nerf: Representing scenes as neural radiance fields for view synthesis,

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoor- thi, and R. Ng, “Nerf: Representing scenes as neural radiance fields for view synthesis,”Communications of the ACM, vol. 65, no. 1, pp. 99–106, 2021

    2021

  7. [15]

    3d gaussian splatting for real-time radiance field rendering

    B. Kerbl, G. Kopanas, T. Leimk ¨uhler, G. Drettakis,et al., “3d gaussian splatting for real-time radiance field rendering.”ACM Trans. Graph., vol. 42, no. 4, pp. 139–1, 2023

    2023

  8. [16]

    Risk of lens cloudiness during colorectal endoscopic submucosal dissection and ability of a novel lens cleaner to maintain and restore endoscopic view,

    N. Yoshida, Y . Naito, R. Hirose, K. Ogiso, K. T. H. Siah, Y . Inada, O. Dohi, K. Kamada, K. Katada, K. Uchiyama,et al., “Risk of lens cloudiness during colorectal endoscopic submucosal dissection and ability of a novel lens cleaner to maintain and restore endoscopic view,”Dig...

    2015

  9. [17]

    A novel lens cleaner to prevent water drop adhesions during colonoscopy and esophagogastroduodenoscopy,

    N. Yoshida, Y . Naito, R. Yasuda, T. Murakami, K. Ogiso, R. Hirose, Y . Inada, O. Dohi, T. Okayama, K. Kamada,et al., “A novel lens cleaner to prevent water drop adhesions during colonoscopy and esophagogastroduodenoscopy,”Endoscopy international open, vol. 5, no. 12, pp. E123...

    2017

  10. [18]

    C3vdv2–colonoscopy 3d video dataset with enhanced realism,

    M. V . Golhar, L. S. G. Fretes, L. Ayers, V . S. Akshintala, T. L. Bobrow, and N. J. Durr, “C3vdv2–colonoscopy 3d video dataset with enhanced realism,”arXiv preprint arXiv:2506.24074, 2025

    2025 arXiv

  11. [19]

    Orb-slam3: An accurate open-source library for visual, visual–inertial, and multimap slam,

    C. Campos, R. Elvira, J. J. G. Rodr ´ıguez, J. M. Montiel, and J. D. Tard ´os, “Orb-slam3: An accurate open-source library for visual, visual–inertial, and multimap slam,”IEEE transactions on robotics, vol. 37, no. 6, pp. 1874–1890, 2021

    2021

  12. [20]

    Cudasift-slam: Multiple-map visual slam for full procedure mapping in real human endoscopy,

    R. Elvira, J. D. Tard ´os, and J. M. M. Montiel, “Cudasift-slam: Multiple-map visual slam for full procedure mapping in real human endoscopy,”arXiv preprint arXiv:2405.16932, 2024

    2024 arXiv

  13. [21]

    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

  14. [22]

    Lightneus: Neural surface reconstruction in endoscopy using illumination de- cline,

    V . M. Batlle, J. M. Montiel, P. Fua, and J. D. Tard ´os, “Lightneus: Neural surface reconstruction in endoscopy using illumination de- cline,” inInternational Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2023, pp. 502–512

    2023

  15. [23]

    3d densifica- tion for multi-map monocular vslam in endoscopy,

    X. Anad ´on, J. Rodr ´ıguez-Puigvert, and J. Montiel, “3d densifica- tion for multi-map monocular vslam in endoscopy,”arXiv preprint arXiv:2503.14346, 2025

    2025 arXiv

  16. [24]

    Mast3r-slam: Real- time dense slam with 3d reconstruction priors,

    R. Murai, E. Dexheimer, and A. J. Davison, “Mast3r-slam: Real- time dense slam with 3d reconstruction priors,” in2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2025, pp. 16 695–16 705

    2025

  17. [25]

    Vggt: Visual geometry grounded transformer,

    J. Wang, M. Chen, N. Karaev, A. Vedaldi, C. Rupprecht, and D. Novotny, “Vggt: Visual geometry grounded transformer,” inPro- ceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 5294–5306

    2025

  18. [26]

    Depth anything 3: Recovering the visual space from any views,

    H. Lin, S. Chen, J. Liew, D. Y . Chen, Z. Li, G. Shi, J. Feng, and B. Kang, “Depth anything 3: Recovering the visual space from any views,”arXiv preprint arXiv:2511.10647, 2025

    2025 arXiv

  19. [27]

    Vggt-slam: Dense rgb slam optimized on the sl (4) manifold,

    D. Maggio, H. Lim, and L. Carlone, “Vggt-slam: Dense rgb slam optimized on the sl (4) manifold,”arXiv preprint arXiv:2505.12549, 2025

    2025 arXiv

  20. [28]

    Direct sparse odometry,

    J. Engel, V . Koltun, and D. Cremers, “Direct sparse odometry,”IEEE transactions on pattern analysis and machine intelligence, vol. 40, no. 3, pp. 611–625, 2017

    2017

  21. [29]

    Nfl-ba: Near-field light bundle adjustment for slam in dynamic lighting,

    A. D. Beltran, D. Rho, M. Niethammer, and R. Sengupta, “Nfl-ba: Near-field light bundle adjustment for slam in dynamic lighting,”arXiv preprint arXiv:2412.13176, 2024

    2024

  22. [30]

    Co-fusion: Real-time segmentation, tracking and fusion of multiple objects,

    M. R ¨unz and L. Agapito, “Co-fusion: Real-time segmentation, tracking and fusion of multiple objects,” in2017 IEEE International Confer- ence on Robotics and Automation (ICRA). IEEE, 2017, pp. 4471– 4478

    2017

  23. [31]

    Dg-slam: Robust dynamic gaussian splatting slam with hybrid pose optimization,

    Y . Xu, H. Jiang, Z. Xiao, J. Feng, and L. Zhang, “Dg-slam: Robust dynamic gaussian splatting slam with hybrid pose optimization,” Advances in Neural Information Processing Systems, vol. 37, pp. 51 577–51 596, 2024

    2024

  24. [32]

    Dgs-slam: Gaussian splatting slam in dynamic environment,

    M. Kong, J. Lee, S. Lee, and E. Kim, “Dgs-slam: Gaussian splatting slam in dynamic environment,”arXiv preprint arXiv:2411.10722, 2024

    2024 arXiv

  25. [33]

    Nr-slam: Nonrigid monocular slam,

    J. J. G. Rodriguez, J. M. Montiel, and J. D. Tardos, “Nr-slam: Nonrigid monocular slam,”IEEE Transactions on Robotics, vol. 40, pp. 4252– 4264, 2024

    2024

  26. [34]

    Free- surgs: Sfm-free 3d gaussian splatting for surgical scene reconstruc- tion,

    J. Guo, J. Wang, D. Kang, W. Dong, W. Wang, and Y .-h. Liu, “Free- surgs: Sfm-free 3d gaussian splatting for surgical scene reconstruc- tion,” inInternational Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2024, pp. 350–360

    2024

  27. [35]

    Endogs: Deformable endoscopic tissues reconstruction with gaussian splat- ting,

    L. Zhu, Z. Wang, J. Cui, Z. Jin, G. Lin, and L. Yu, “Endogs: Deformable endoscopic tissues reconstruction with gaussian splat- ting,” inInternational Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2024, pp. 135–145

    2024

  28. [36]

    Endo-4dgs: Endoscopic monocular scene reconstruction with 4d gaussian splatting,

    Y . Huang, B. Cui, L. Bai, Z. Guo, M. Xu, M. Islam, and H. Ren, “Endo-4dgs: Endoscopic monocular scene reconstruction with 4d gaussian splatting,” inInternational Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2024, pp. 197–207

    2024

  29. [37]

    Nice-slam: Neural implicit scalable encoding for slam,

    Z. Zhu, S. Peng, V . Larsson, W. Xu, H. Bao, Z. Cui, M. R. Oswald, and M. Pollefeys, “Nice-slam: Neural implicit scalable encoding for slam,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 12 786–12 796

    2022

  30. [38]

    Point- slam: Dense neural point cloud-based slam,

    E. Sandstr ¨om, Y . Li, L. Van Gool, and M. R. Oswald, “Point- slam: Dense neural point cloud-based slam,” inProceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 18 433–18 444

    2023

  31. [39]

    Gaussian splatting slam,

    H. Matsuki, R. Murai, P. H. Kelly, and A. J. Davison, “Gaussian splatting slam,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 18 039–18 048

    2024

  32. [40]

    Extrinsplat: Decoupling geometry and semantics for open-vocabulary understanding in 3d gaussian splatting,

    J. Ding, X. Liu, Z. Pan, S. Long, and G. Li, “Extrinsplat: Decoupling geometry and semantics for open-vocabulary understanding in 3d gaussian splatting,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 31 019–31 028

    2026

  33. [41]

    Desplat: Decomposed gaussian splatting for distractor- free rendering,

    Y . Wang, M. Klasson, M. Turkulainen, S. Wang, J. Kannala, and A. Solin, “Desplat: Decomposed gaussian splatting for distractor- free rendering,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 722–732

    2025

  34. [42]

    Hybridgs: Decoupling transients and statics with 2d and 3d gaussian splatting,

    J. Lin, J. Gu, L. Fan, B. Wu, Y . Lou, R. Chen, L. Liu, and J. Ye, “Hybridgs: Decoupling transients and statics with 2d and 3d gaussian splatting,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 788–797

    2025

  35. [43]

    Grounding image matching in 3d with mast3r,

    V . Leroy, Y . Cabon, and J. Revaud, “Grounding image matching in 3d with mast3r,” inEuropean conference on computer vision. Springer, 2024, pp. 71–91

    2024

  36. [44]

    A benchmark for the evaluation of rgb-d slam systems,

    J. Sturm, N. Engelhard, F. Endres, W. Burgard, and D. Cremers, “A benchmark for the evaluation of rgb-d slam systems,” in2012 IEEE/RSJ international conference on intelligent robots and systems. IEEE, 2012, pp. 573–580

    2012

  37. [45]

    Image quality assessment: from error visibility to structural similarity,

    Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: from error visibility to structural similarity,”IEEE transactions on image processing, vol. 13, no. 4, pp. 600–612, 2004

    2004

  38. [46]

    The unreasonable effectiveness of deep features as a perceptual metric,

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang, “The unreasonable effectiveness of deep features as a perceptual metric,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 586–595

    2018

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.