{"id":"49df561b-8e4b-41d9-945e-9fa464c3d880","arxiv_id":"2501.01101","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"EH-SurGS combines additive opacity life cycles and an adaptive static/dynamic region mask to achieve higher PSNR and faster rendering than previous deformable surgical scene reconstruction methods.","lead":"This paper introduces EH-SurGS, a 3D Gaussian splatting method for reconstructing moving surgical scenes from endoscope video. It adds a life cycle model so scene elements can appear and disappear, and a region mask that skips deformation for static areas, reporting better image quality and faster rendering than prior methods.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Adaptive motion hierarchy speedup depends on an unspecified image-region-to-Gaussian assignment; without it the reported FPS gain over Deform3DGS (~5%) is not reproducible.","rationale":"The paper's strongest empirical claim is that EH-SurGS outperforms state-of-the-art methods in both reconstruction quality and rendering speed. The quality improvement is attributed mainly to the life-cycle deformation module, while the speed improvement is attributed specifically to the adaptive motion hierarchy strategy (AMHS). The ablation in Table II shows that removing AMHS leaves PSNR, SSIM, and LPIPS essentially unchanged (39.92 vs 39.91 PSNR, 0.972 vs 0.972 SSIM, 0.034 vs 0.034 LPIPS) and only reduces FPS from 379.67 to 351.00. Thus the entire empirical support for the AMHS contribution is an FPS gain of about 8%. That gain is modest compared with the 361.33 FPS already achieved by Deform3DGS, which deforms all Gaussians. Whether the gain is real depends on how image-space mask regions are mapped to individual 3D Gaussians, and the paper leaves this unspecified. Because the projection of a 3D Gaussian is a 2D footprint, not a point, and because densification and deformation continuously change the set and locations of Gaussians, there is no canonical way to infer the intended mapping. Different plausible mappings could skip different fractions of Gaussians and could introduce boundary artifacts when a Gaussian straddles a static/dynamic boundary. The paper reports neither the fraction of Gaussians classified as static nor any sensitivity to the hand-set thresholds δ1 and δ2. Without the mapping and without code, the speed claim cannot be independently verified. This is a correctness and reproducibility risk, not merely a stylistic omission. The missing EndoSparse baseline and the LPIPS regression on StereoMIS P1-1 also weaken the blanket SOTA statement, but they are comparability issues rather than threats to the internal mechanism. The reader's weakest_assumption identified the same mapping problem, and my analysis agrees that it is the load-bearing point. Since the concern does not prove the method wrong but does show that the central speed claim is currently under-specified, the appropriate verdict remains conditional.","tokens_in":11084,"tokens_out":9423,"duration_ms":100681,"concrete_test":"Release code with the exact region-to-Gaussian assignment, or independently reimplement it using the most natural rule: project each Gaussian's current mean into the image and mark it static if its cell is static. Rerun EndoNeRF-Cutting and compare PSNR/SSIM/LPIPS/FPS and the per-frame fraction of skipped Gaussians with Table II. Then repeat with two alternative assignment rules (full 2D footprint overlap; initial depth back-projection) and with δ1, δ2 varied by ±50%. If any variant drops FPS below Deform3DGS's 361.33 or changes PSNR by more than the reported variance, the speedup claim is not robust. If the authors provide code, the same test is a direct reproducibility check rather than an inference.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central speed claim (Sec. IV-A) rests on the adaptive motion hierarchy strategy of Sec. III-C. The mask F is defined in image space (H×W, split into N×N regions), but the method must decide, for every 3D Gaussian, whether it lies in a static or deformable region. The paper never defines this correspondence. The update criteria in Sec. III-C compute 'average deformation' and losses 'for each region,' yet a Gaussian is a 3D object whose projection can overlap several regions, whose center can lie in a different region than its footprint, and whose assignment can change as it deforms and as densification adds new Gaussians. Possible rules (projected mean, 2D footprint overlap, initial depth back-projection) are mutually inconsistent after optimization, and the text gives no pseudocode or released code. This is not a cosmetic omission: the ablation in Table II shows the entire benefit of AMHS is an FPS increase from 351.00 to 379.67, i.e., about 8%; Deform3DGS, which deforms all Gaussians, already runs at 361.33. If a reasonable mapping skips fewer Gaussians or causes boundary artifacts, EH-SurGS could be slower or less accurate than a baseline that is claimed to be outperformed. The thresholds δ1=0.05 and δ2=0.5 and the update rule Eq. (9) are additionally fixed without sensitivity analysis, but the mapping is the primary unstated assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes EH-SurGS, a deformable 3D Gaussian splatting method for reconstructing surgical scenes from stereo endoscopic video. It introduces a life-cycle model for 3D Gaussians to handle irreversible tissue changes and an adaptive motion hierarchy strategy (AMHS) that classifies image regions as static or deformable so that only dynamic Gaussians pass through the deformation field. The method is evaluated on six clips from the EndoNeRF, Hamlyn, and StereoMIS datasets, reporting higher PSNR and SSIM than four 3DGS-based baselines and Forplane, higher FPS than Deform3DGS, and a lower LPIPS on most clips. Ablations on EndoNeRF-Cutting isolate the contributions of the life cycle and the motion hierarchy.","tokens_in":11352,"tokens_out":12328,"duration_ms":104242,"significance":"The ideas are relevant to real-time surgical scene reconstruction, and the reported improvements are consistent across datasets. The life-cycle formulation and the static/deformable decomposition are reasonable and, if fully specified, would be a useful contribution. The paper also provides a thorough ablation, including a comparison of additive versus multiplicative life-cycle opacities. However, the manuscript currently omits the key detail that makes AMHS workable—the assignment from image-space regions to 3D Gaussians—and does not address validity constraints on deformed Gaussian parameters. Because the code is promised but not yet available, these omissions materially block reproduction.","major_comments":[{"comment":"The adaptive motion hierarchy strategy is under-specified: the mask F is defined in image space (H×W, divided into N×N regions), but the paper never states how a region label is assigned to each 3D Gaussian. The text says 'The updated mask will control whether the 3D Gaussians in the canonical space need to be deformed' and computes average deformation 'for all 3D Gaussians in each region,' yet a Gaussian can project onto multiple regions, can move between regions as it deforms, and can be newly created by densification. Without an explicit assignment rule (projected center, footprint overlap, depth back-projection, or similar) and a rule for newly added Gaussians, the speedup of AMHS reported in Table II (351.00 to 379.67 FPS) is not reproducible. Please provide the full assignment procedure, including behavior under the region-splitting step.","section":"Section III-C, Fig. 2"},{"comment":"The deformation model applies unconstrained additive updates to rotation, scale, and opacity. In 3DGS, rotation is a unit quaternion, but adding an unconstrained increment in Eq. (7) does not preserve unit length; scale values can become negative; and the additive opacity update in Eq. (8) can produce alpha_t below 0 or above 1. These would make the alpha compositing in Eqs. (4)-(5) invalid. The paper does not describe any normalization, clamping, or reparameterization for these parameters. Since the reported quality gain of the life-cycle model (Table II) depends on this update, please clarify how valid parameter ranges are enforced during optimization.","section":"Section III-B, Eqs. (7)-(8)"},{"comment":"The baseline list in Section IV includes EndoSparse [11], but Table I contains no result for EndoSparse on any clip. Because the paper claims superiority over state-of-the-art methods, omitting a named baseline from the quantitative comparison leaves the claim incomplete. Please add the missing results or explicitly justify the exclusion.","section":"Section IV, Baselines and Table I"},{"comment":"The sentence 'EH-SurGS surpasses these methods in terms of reconstruction quality, training time, and rendering speed across all six datasets' is contradicted by Table I on StereoMIS-P1-2: LGS reports a training time of 119 s versus 131 s for EH-SurGS. Please correct the claim to reflect the actual training-time comparisons.","section":"Section IV-A, Table I"},{"comment":"The thresholds δ1=0.05 and δ2=0.5 are described as 'experimentally determined,' but the paper provides no sensitivity analysis or validation procedure. If these constants were tuned on the same clips used for final evaluation, the results in Table I may be optimistic. Please document how the thresholds were chosen and report a sensitivity study over their values, or provide a cross-validation protocol.","section":"Section III-C, update criteria"},{"comment":"The statement that 'the projection matrix P is set as the identity matrix' is difficult to interpret, because a 4×4 identity matrix is not a valid perspective projection for an endoscopic camera. If P is instead intended to be the camera pose or a pre-warped mapping, this should be stated precisely; otherwise the rendering procedure cannot be reproduced from the paper.","section":"Section III-D, Implementation Details"}],"minor_comments":[{"comment":"The norms in LC and LD are not specified; please state whether L1 or L2 is used, and define the ranking loss L_rank precisely.","section":"Eq. (10)"},{"comment":"The parentheses are said to contain variance, but the magnitudes (e.g., 0.011 for PSNR) look like standard deviations; please clarify the statistic and report the number of runs.","section":"Table I caption"},{"comment":"The datasets section cites the StereoMIS dataset as [20], but Section IV-A cites '[25]', and reference [25] is an unrelated room-reconstruction paper; please correct the citation.","section":"Section IV, Datasets"},{"comment":"The w/o MIU variant 'maintains the default number of mask updates,' but the initial value of Nm and the update schedule are not defined in Section III-C; please specify them.","section":"Section IV-B, Ablation"},{"comment":"The training/test split is described only as a '7:1 ratio'; please report the exact number of training and testing frames for each clip, since the clip lengths vary from 63 to 184 frames.","section":"Section IV, Datasets"},{"comment":"The figure would benefit from explicitly showing how the image-space mask is converted into per-Gaussian static/deformable labels, since this conversion is a central step of the method.","section":"Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The strongest concern is the missing mapping between the image-space mask and the 3D Gaussians; without it, the speed claim cannot be assessed. The paper also has a citation inconsistency and an over-broad claim about training time. If the authors can provide the missing details and correct the errors, the work could be acceptable after revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: EH-SurGS is a sound, incremental engineering paper for deformable surgical scene reconstruction. The two contributions — additive life-cycle opacity and an adaptive static/dynamic region mask — are small variations of existing ideas, but they work in the sense that the ablations show each helps somewhat on the EndoNeRF-Cutting clip. The experiments are reasonably thorough: three datasets, six clips, three runs per setting, and consistent (if modest) PSNR/SSIM gains over Deform3DGS, on the order of 0.2-1.5 dB. The paper is clearly written, and the problem it targets — irreversible tissue shearing — is real.\n\nThe soft spots are real too, and one of them is load-bearing. The speedup claim rests on the adaptive motion hierarchy strategy, but the paper never defines how an image-space region mask (H×W divided into N×N blocks) is mapped to individual 3D Gaussians. The text says we compute average deformation and losses 'for each region' and then the mask controls which Gaussians pass through the deformation field, but a Gaussian is a 3D ellipsoid whose footprint can span several regions; the assignment rule is never given. Without that rule the FPS numbers — 379 vs 361 for Deform3DGS on Cutting, similar elsewhere — are not reproducible. The ablation shows the whole mask mechanism buys about 8% FPS on Cutting, which is a thin margin for the added complexity.\n\nSeveral lesser issues: δ1 and δ2 are hand-set with no sensitivity analysis; EndoSparse is named as a baseline in Sec. IV but is absent from Table I; on StereoMIS P1-1 the LPIPS is worse than Deform3DGS (0.177 vs 0.170), which slightly contradicts the blanket 'significant superiority' claim. No code is released.\n\nVerdict: this deserves a serious referee — the problem is genuine and the empirical work is honest — but it needs major revision before acceptance. The authors should specify the Gaussian-to-region mapping precisely (projected mean, footprint overlap, etc.), add sensitivity analysis for the thresholds, include EndoSparse, and address the LPIPS regression. If the code were released, many of these concerns would be answerable directly.","headline":"EH-SurGS is a competent incremental contribution whose quality gains are real but modest, and whose speed claim is not reproducible because the key image-region-to-Gaussian mapping is never specified.","tokens_in":11895,"tokens_out":5559,"would_cite":false,"duration_ms":44857,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Surgical scene rendering hits up to 400 fps with sharper tissue detail","keywords":["3D Gaussian splatting","deformable scene reconstruction","surgical scene reconstruction","tissue shearing","life cycle modeling","adaptive motion hierarchy","endoscopic video","real-time rendering"],"falsifier":"Train EH-SurGS on a surgical clip with a static background and one shearing event, then render the same frames with the adaptive motion hierarchy disabled: the paper predicts equal reconstruction quality but lower FPS. If enabling the hierarchy lowers PSNR or introduces ghosting around the sheared region, the mask is misclassifying deformable Gaussians and the central speed-quality claim fails. A second check: inspect depth maps in sheared regions; the additive life-cycle opacity should remove old Gaussians cleanly, whereas multiplicative opacity leaves depth artifacts, as the paper's ablations show.","tokens_in":10876,"feed_emoji":"🏥","tokens_out":5375,"duration_ms":44629,"temperature":0.7,"pith_summary":"This paper proposes EH-SurGS, a 3D Gaussian splatting method for reconstructing deformable surgical scenes from stereo endoscopic video. It claims that adding a learnable 'life cycle' to each Gaussian—so opacity can rise and fall over time—lets the model capture irreversible tissue changes such as cutting and shearing, not just reversible motion. It further claims an adaptive motion hierarchy that classifies image regions as static or deformable can skip the deformation field for static Gaussians, raising rendering speed without hurting quality. On six clips from the EndoNeRF, Hamlyn, and StereoMIS datasets, the reported results show higher PSNR and SSIM than all compared baselines and rendering speeds of roughly 349–400 frames per second. If the claims hold, surgical scene reconstruction becomes fast enough for real-time use while better representing tissue damage.","feed_headline":"Surgical scene rendering hits up to 400 fps with sharper tissue detail","feed_subtitle":"A life-cycle model captures tissue shearing while an adaptive mask skips static Gaussians during deformation.","key_machinery":"The two load-bearing components are the life-cycle opacity model and the adaptive motion mask. The life cycle makes opacity itself time-dependent through an additive sum of learnable Gaussian basis functions (Eq. 8), which the paper distinguishes from prior multiplicative opacity products; this is what lets it represent newly appearing or disappearing content such as sheared tissue. The motion mask is an $H\\times W$ image-space label map, updated every $N_m$ iterations, that decides per region whether Gaussians must be warped; thresholds $\\delta_1$ and $\\delta_2$ plus recursive splitting of conflicting regions form the hierarchy. Together they determine which Gaussians enter the deformation field and at what opacity they render.","core_discovery":"EH-SurGS extends deformable 3D Gaussian splatting with two mechanisms. First, for irreversible deformations, each Gaussian's opacity is modeled as $\\alpha_t = \\alpha_0 + \\sum_{j=1}^{B} \\omega^\\alpha_j b^\\alpha(t)$, with learnable Gaussian basis functions $b(t)=\\exp\\left(-\\frac{1}{2\\sigma^2}(t-\\theta)^2\\right)$ and $B=20$, so Gaussians can appear and disappear at specific times instead of only warping. Second, an adaptive motion hierarchy divides the image into $N\\times N$ regions and maintains a mask $F$ that labels regions as dynamic or static using two criteria: average Gaussian displacement against threshold $\\delta_1=0.05$, and rendering-loss difference with and without deformation against $\\delta_2=0.5$; conflict regions are split into four subregions. Gaussians in static regions bypass the deformation field. The paper reports that on EndoNeRF-Cutting/Pulling, Hamlyn-Clip1/2, and StereoMIS-P1-1/P1-2, EH-SurGS outperforms Forplane, Deform3DGS, Endo-4DGS, LGS, and SurgicalGaussian in PSNR and SSIM, with FPS up to 400, while the ablations attribute quality gains to the life cycle and speed gains to the motion hierarchy.","pith_inferences":["The image-space mask implies an unstated geometric correspondence between image regions and 3D Gaussians; making the mask operate directly on 3D Gaussians (e.g., by spatial proximity) would be a natural follow-up that could extend the speedup to moving-camera settings.","The same life-cycle mechanism could transfer to other dynamic scene reconstruction tasks where objects appear or disappear—such as removing surgical instruments or modeling cutting—beyond the endoscopic domain.","Since the paper is limited to static endoscopes, a direct test of the hierarchy is to keep the camera fixed but add parallax-inducing motion; if the mask mislabels background Gaussians under small camera motion, the speedup will not survive extension to colonoscopy.","The reported FPS counts assume the mask is computed infrequently; an implementation that updates the mask every frame would likely trade the speed advantage away, so the practical gain depends on the update schedule."],"forward_implications":["Real-time rendering: the reported 349–400 FPS means deformable surgical scenes can be rendered at interactive or real-time rates on a single RTX 4090.","Better fidelity for tissue damage: the life cycle represents irreversible changes like cutting and shearing, which prior deformation-only models blur or ghost.","Faster training: EH-SurGS reports lower training time than Forplane, Endo-4DGS, LGS, and SurgicalGaussian, with training times around 97–131 seconds across the six clips.","Modular components: the ablation shows the life cycle and motion hierarchy are independent; removing either one degrades quality or speed respectively.","The ranking loss from SparseNeRF is retained for depth supervision, indicating the method does not sacrifice geometric constraints for speed."],"supporting_citations":[{"why":"Provides the 3D Gaussian splatting representation and differentiable tile rasterizer that the canonical space and renderer are built on.","marker":"[14]"},{"why":"Deform3DGS is the main deformable-Gaussian baseline and the source of the dense point-cloud initialization procedure.","marker":"[9]"},{"why":"Endo-4DGS is a 3DGS-based deformable surgical baseline that EH-SurGS must beat in quality and speed.","marker":"[10]"},{"why":"Supplies the EndoNeRF dataset (Pulling and Cutting clips) used for evaluation and ablation.","marker":"[4]"},{"why":"Introduces the multiplicative temporal-lifecycle product that the paper contrasts with its additive opacity model.","marker":"[17]"},{"why":"GaussianPrediction is the other multiplicative-opacity approach the paper contrasts with its life-cycle formulation.","marker":"[18]"},{"why":"Supplies the depth ranking loss $L_{\\text{rank}}$ incorporated into the full objective.","marker":"[19]"},{"why":"Provides the StereoMIS dataset clips used to evaluate natural and artificial deformations.","marker":"[20]"},{"why":"Provides the Hamlyn dataset clips used for evaluation, with rectified images, stereo depth, and camera calibration.","marker":"[21]"},{"why":"Generates surgical instrument masks for the Hamlyn clips.","marker":"[22]"}],"fun_headline_variants":["EH-SurGS: life-cycle Gaussians handle shearing at 400 fps","Surgical scene rendering at 400 fps with a life-cycle model","Adaptive motion hierarchy boosts surgical Gaussian splatting to 400 fps","Life-cycle Gaussians and adaptive masks make surgical scenes render at 400 fps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that an image-space region mask, updated with thresholds on average deformation and loss difference, correctly tells which 3D Gaussians are static, so those Gaussians can skip deformation without visible error.","fun_headline_variants_meta":{"raw":{"variants":["EH-SurGS: life-cycle Gaussians handle shearing at 400 fps","Surgical scene rendering at 400 fps with a life-cycle model","Adaptive motion hierarchy boosts surgical Gaussian splatting to 400 fps","Life-cycle Gaussians and adaptive masks make surgical scenes render at 400 fps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001778,"raw_usage":{"total_tokens":7068,"prompt_tokens":1059,"completion_tokens":6009,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":675,"completion_tokens_details":{"reasoning_tokens":5928}},"tokens_in":675,"tokens_out":6009,"duration_ms":39073,"temperature":1.0,"reasoning_tokens":5928,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:34:42.667502+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train EH-SurGS on a surgical clip with a static background and one shearing event, then render the same frames with the adaptive motion hierarchy disabled: the paper predicts equal reconstruction quality but lower FPS. If enabling the hierarchy lowers PSNR or introduces ghosting around the sheared region, the mask is misclassifying deformable Gaussians and the central speed-quality claim fails. A second check: inspect depth maps in sheared regions; the additive life-cycle opacity should remove old Gaussians cleanly, whereas multiplicative opacity leaves depth artifacts, as the paper's ablations show.","supporting_citations":[{"cited_title":"Neural rendering for stereo 3d reconstruction of deformable tissues in robotic surgery,","cited_arxiv_id":null,"evidence_quote":"Supplies the EndoNeRF dataset (Pulling and Cutting clips) used for evaluation and ablation."},{"cited_title":"Gaussianprediction: Dynamic 3d gaussian prediction for motion extrapolation and free view synthesis,","cited_arxiv_id":null,"evidence_quote":"GaussianPrediction is the other multiplicative-opacity approach the paper contrasts with its life-cycle formulation."},{"cited_title":"Sparsenerf: Distilling depth ranking for few-shot novel view synthesis,","cited_arxiv_id":null,"evidence_quote":"Supplies the depth ranking loss $L_{\\text{rank}}$ incorporated into the full objective."},{"cited_title":"Learning how to robustly estimate camera pose in endoscopic videos,","cited_arxiv_id":null,"evidence_quote":"Provides the StereoMIS dataset clips used to evaluate natural and artificial deformations."},{"cited_title":"Endo- depth-and-motion: Reconstruction and tracking in endoscopic videos using depth networks and photometric constraints,","cited_arxiv_id":null,"evidence_quote":"Provides the Hamlyn dataset clips used for evaluation, with rectified images, stereo depth, and camera calibration."},{"cited_title":"Efficientsam: Leveraged masked image pretraining for efficient segment anything,","cited_arxiv_id":null,"evidence_quote":"Generates surgical instrument masks for the Hamlyn clips."}],"review_version":1}