{"id":"0532d143-4bef-4af2-b0fc-f1a97f43f681","arxiv_id":"2608.08949","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":11,"one_line_summary":"A Gaussian Splatting SLAM system that gates map updates with a reliability gate and separates camera-attached artifacts into a transient field, improving endoscopic tracking and reconstruction under optical degradation.","lead":"EndoMD-SLAM adds a memory bank and a static-transient decomposition to Gaussian Splatting SLAM so colonoscopy videos with water droplets or debris on the lens no longer corrupt the 3D map. On a curated colonoscopy benchmark it reports 91% lower trajectory error and 9.9 dB better rendering than the weakest baseline, though no code or error bars are provided.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Tracking depends entirely on MASt3R correspondences after transient masking, but the paper reports no evidence of correspondence density or mask accuracy in textureless colon, so the robustness claim may not generalize beyond the ten curated sequences.","rationale":"The paper is well-motivated and the full system performs strongly on the curated benchmark, but the central claim depends on an unvalidated assumption about MASt3R reliability in textureless colon images after transient masking. The reader identified this same weakest point. The proposed concrete check would settle it. I also note the ablation table's baseline row matching EndoGSLAM exactly is a genuine inconsistency that should be clarified, but it does not by itself invalidate the headline benchmark results. The CONDITIONAL verdict remains appropriate: the method is promising, but needs released code, per-sequence statistics, and the proposed correspondence-density analysis before the central claim can be accepted with confidence.","tokens_in":12653,"tokens_out":10998,"duration_ms":103058,"concrete_test":"Run MASt3R on the ten C3VDv2 sequences in both clean and degraded frames; for each frame pair, compute the weighted Procrustes inlier count N_valid (Eq. 2) and match fraction rho before and after applying the trained transient mask, and report the per-sequence distributions against the gate thresholds tau_rho=0.15 and tau_n=500. If a non-negligible fraction of frames fall below threshold even after masking, and relocalization's top hypothesis also fails, the tracking assumption is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central robustness claim rests on the tracking chain in Section III-A: MASt3R must supply enough metric 3D correspondences in textureless colon images, and the per-frame transient opacity mask from Section III-B must suppress the high-confidence matches that MASt3R assigns to moving lens contaminants (Eq. 1). The paper gives no quantitative evidence that this holds. No correspondence density, inlier fraction, or mask-quality metric is reported; the reliability gate thresholds (tau_rho=0.15, tau_n=500) are asserted without showing that typical colon frames clear them. The transient mask is learned from poses produced by the same tracking loop, so a wrong pose corrupts the mask and vice versa, creating a circular dependency. Table II is also internally suspicious: the 'Baseline' row is numerically identical to EndoGSLAM in Table I even though EndoMD-SLAM is described as using a different foundation-model tracker; this blurs whether the ablation isolates the proposed components on the proposed architecture. If MASt3R fails on clean or mildly degraded colon tissue, the memory gate and relocalization (Eqs. 3-4) cannot restore tracking, and the benchmark success would not transfer to routine colonoscopy.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":12912,"tokens_out":5485,"duration_ms":54765,"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":[{"comment":"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.","section":"Section IV-C, Table II"},{"comment":"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.","section":"Section IV-A-d, Table I"},{"comment":"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.","section":"Section III-A-a, Eqs. (1)-(3)"},{"comment":"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.","section":"Section III-A-a and III-B"}],"minor_comments":[{"comment":"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.","section":"Abstract, Section IV-B"},{"comment":"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.","section":"Section IV-A-d"},{"comment":"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.","section":"Section III-B, Eq. (7)-(8)"},{"comment":"There is a typo in Section IV-D: 'estimatsed' should be 'estimated'.","section":"Section IV-D"},{"comment":"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.","section":"Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope for a computer vision venue, but the current evidence does not yet establish the central robustness claim. The most concerning point is not the method's motivation but the consistency of the experimental evaluation: the ablation baseline matching EndoGSLAM exactly, the absence of per-sequence statistics, and the lack of any validation procedure for the many thresholds. I would encourage the editor to ask for a revised version with a corrected ablation and additional statistical and sensitivity analyses before considering acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is worth a careful read. The problem is real: in endoscopic procedures, water flushing and debris on the lens break the multi-view photometric consistency that Gaussian Splatting SLAM relies on, causing tracking drift and permanent map corruption. The proposed solution is well-motivated: a reliability gate that suspends map updates during unreliable frames, a memory bank of clean keyframes for drift-aware relocalization, and a per-frame transient Gaussian field that absorbs camera-attached contaminants. The reported gains are large (ATE from ~36 mm to ~3 mm, PSNR from ~8 dB to ~18 dB on a curated subset of C3VDv2), and if they hold, this is a solid step for clinical 3D reconstruction.\n\nWhat is genuinely new is the combination: MASt3R foundation-model tracking re-weighted by the transient opacity mask, coupled with a memory bank and static-transient decomposition for online SLAM. The paper also contributes a degradation-focused benchmark, which the community will find useful. The writing is clear and the related work is placed honestly.\n\nThat said, the evaluation has real soft spots. First, the benchmark is ten curated sequences from a synthetic dataset, with no per-sequence statistics or error bars. The reliability-gate thresholds and loss weights are asserted without a validation split, so the numbers may partly reflect tuning on the test set. Second, the ablation table is internally suspicious: the \"Baseline\" row is numerically identical to EndoGSLAM in the main table even though EndoMD-SLAM uses a different tracker. That blurs what the ablation actually isolates—the memory-only row being worse than baseline makes it particularly hard to interpret. Third, the tracking chain leans entirely on MASt3R correspondences after transient-mask re-weighting, but there is no evidence about mask accuracy or correspondence density in textureless colon tissue. The reliability gate is a plausible safety net, but if MASt3R fails on clean or mildly degraded frames, the system may not recover.\n\nThis paper is for researchers working on robust 3DGS SLAM and medical 3D reconstruction. It deserves a serious referee: the idea is sound, the problem matters, and the results are promising. The concerns are fixable with per-sequence statistics, a held-out split or independent dataset, a cleaner ablation, and some analysis of mask/correspondence quality. I would send it to review.","headline":"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.","tokens_in":13515,"tokens_out":4652,"would_cite":true,"duration_ms":44829,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"EndoMD-SLAM keeps colonoscopy mapping stable under lens contamination by gating map updates and separating transient artifacts from the anatomical field.","keywords":["endoscopic SLAM","Gaussian splatting","optical degradation","static-transient decomposition","temporal memory","relocalization","colonoscopy 3D reconstruction","RGB-D SLAM"],"falsifier":"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.","tokens_in":12404,"feed_emoji":"🩺","tokens_out":11496,"duration_ms":109655,"temperature":0.7,"pith_summary":"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.","feed_headline":"Colonoscopy mapping that ignores lens grime cuts trajectory error 91%","feed_subtitle":"The system pauses map updates during flushing and stores droplets in a separate field, keeping the 3D anatomy clean.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Supplies the dense 3D correspondences and descriptors used for transient-aware tracking, pose solving, and memory-bank relocalization.","marker":"[43]"},{"why":"Provides the C3VDv2 colonoscopy dataset from which the ten degraded sequences are filtered.","marker":"[18]"},{"why":"The endoscopy-specific 3DGS SLAM baseline whose failure motivates the design and against which the headline ATE and PSNR gains are reported.","marker":"[4]"},{"why":"General-purpose 3DGS SLAM baseline included in the benchmark comparison.","marker":"[39]"},{"why":"NeRF-based SLAM baseline representing the implicit-reconstruction approach in the comparison.","marker":"[37]"},{"why":"Defines the explicit 3D Gaussian representation on which both the static and transient fields are built.","marker":"[15]"},{"why":"Supplies the endoscopy-specific evaluation conventions and RGB-D input setting that the experiments adopt.","marker":"[5]"},{"why":"Offline distractor/transient decomposition formulation that motivates separating static scene content from image-specific transients.","marker":"[41]"},{"why":"Defines the ATE and RMSE evaluation metrics used to report tracking and geometry accuracy.","marker":"[44]"}],"fun_headline_variants":["Splitting droplets from anatomy keeps colonoscopy maps stable","Pausing map updates during flushing cuts trajectory error 91%","Separating transient dirt from static anatomy reduces drift by 91%","Memory-gated SLAM ignores lens grime, cuts ATE by 91%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Splitting droplets from anatomy keeps colonoscopy maps stable","Pausing map updates during flushing cuts trajectory error 91%","Separating transient dirt from static anatomy reduces drift by 91%","Memory-gated SLAM ignores lens grime, cuts ATE by 91%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000466,"raw_usage":{"total_tokens":2354,"prompt_tokens":1004,"completion_tokens":1350,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":620,"completion_tokens_details":{"reasoning_tokens":1274}},"tokens_in":620,"tokens_out":1350,"duration_ms":12815,"temperature":1.0,"reasoning_tokens":1274,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:18:52.721034+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"Endogslam: Real-time dense reconstruction and tracking in endoscopic surgeries using gaussian splatting,","cited_arxiv_id":null,"evidence_quote":"The endoscopy-specific 3DGS SLAM baseline whose failure motivates the design and against which the headline ATE and PSNR gains are reported."},{"cited_title":"Endoflow-slam: Real-time endoscopic slam with flow-constrained gaussian splatting,","cited_arxiv_id":null,"evidence_quote":"Supplies the endoscopy-specific evaluation conventions and RGB-D input setting that the experiments adopt."},{"cited_title":"Desplat: Decomposed gaussian splatting for distractor- free rendering,","cited_arxiv_id":null,"evidence_quote":"Offline distractor/transient decomposition formulation that motivates separating static scene content from image-specific transients."}],"review_version":1}