{"id":"7228d25a-c5c3-41a4-b5cf-2b04ceedbfa6","arxiv_id":"2501.01465","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A modular endoscopic 3D reconstruction pipeline is described, with Depth-Anything V2 and EndoDAC for depth and adaptive-threshold ICP for merging, but predicted-depth reconstructions are still qualitatively poor.","lead":"This tech report presents a modular pipeline that selects frames, estimates depth with two learning-based models, and merges depth maps into 3D point clouds for endoscopic surgery. It reports that Depth-Anything V2 outperforms EndoDAC on Hamlyn depth metrics, but that reconstructed point clouds from predicted depths remain far from ground truth.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The report's own Section 4.3 states DA:v2 reconstructions cannot be aligned to ground truth and are 'far worse', directly contradicting the abstract's integrated-effectiveness claim; this is the load-bearing weakness, ahead of the per-sequence depth scaling.","rationale":"The reader's conditional verdict is reasonable, but the weakest assumption they name (retrospective scaling in §4.1) is not the first place the central claim breaks. Even granting the depth-ranking, the abstract's claim that the integrated methods are effective is contradicted by §4.3, which states that reconstruction from DA:v2 cannot be aligned to ground truth, the discrepancy is too large, and the depth-estimation images are 'far worse' than ground-truth reconstructions. §3 further admits the frame-selection module was not tested. Thus no end-to-end validation exists for the advertised pipeline. The Section 4.3 admission is an explicit, located limitation that must be weighed, and it lands: the 'usable baseline for endoscopic 3D reconstruction' component of the strongest claim is not supported. A quantitative reconstruction experiment with aligned poses and a global similarity transform would settle whether the failure is simply missing scale/pose or a deeper property of monocular depth; until that is reported, the abstract should be tempered. This does not require rejecting the report—its transparency about limitations supports conditional acceptance pending revision.","tokens_in":9313,"tokens_out":9835,"duration_ms":98573,"concrete_test":"Run the full pipeline end-to-end on the first 91 frames of test22 using DA:v2, neighbor ICP with the recommended mean+2σ threshold, and a pose estimate from Hamlyn ground truth (if available) or a standard VO/SfM trajectory; apply a single global Umeyama alignment of the final point cloud to the Hamlyn ground-truth surface and report Chamfer/RMS point-to-surface error. Compare with the same reconstruction from ground-truth depth images. If the DA:v2 error is substantially larger and remains above a clinically relevant threshold, Section 4.3's admission is confirmed and the integrated-effectiveness claim must be withdrawn or explicitly limited.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the integrated pipeline—frame selection, DA:v2 depth, neighbor ICP—yields a usable 3D reconstruction. The report's own Section 4.3 undermines this: it says a quantitative reconstruction comparison 'is not feasible at this point', the discrepancy between the reconstructed scene and the actual scene is 'too large', a single scaling cannot align the point clouds, and the DA:v2 result in Figure 13 is 'far worse' than a ground-truth-depth reconstruction. Section 3 also states the frame-selection module 'was not tested', so no end-to-end validation of the advertised integration exists. This is an internal contradiction with the abstract's 'Experiments ... demonstrate the effectiveness of the integrated methods.' The per-sequence scaling fitted for the depth tables in Section 4.1 is a valid secondary concern for the DA:v2-vs-others ranking, but it is not the most load-bearing issue: even if DA:v2 is the best depth estimator on Hamlyn, the paper itself reports that the reconstruction stage fails to produce an aligned, usable model. The report should either supply quantitative reconstruction error or downgrade the abstract to a component-level depth-estimation claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a modular, open-source pipeline for endoscopic 3D reconstruction from monocular video, combining frame selection (HyperIQA and R-channel intensity), depth estimation (Depth-Anything V1/V2 and EndoDAC), and ICP-based point-cloud alignment with a dynamic thresholding mechanism. The experimental sections report quantitative depth estimation comparisons on two Hamlyn sequences (test22 and rectified01), a search over seven ICP threshold schemes on a single frame pair, and a qualitative assessment of end-to-end reconstruction. The paper's stated contribution is a flexible pipeline architecture plus evidence that Depth-Anything V2 outperforms EndoDAC and Depth-Anything V1 for the tested depth metrics, with the abstract additionally claiming that the integrated methods are effective.","tokens_in":9583,"tokens_out":4782,"duration_ms":50150,"significance":"If the claims were supported, the paper would provide a useful modular baseline and open-source implementation for endoscopic reconstruction, with a candid account of practical difficulties such as relative depth scaling and ground-truth inconsistencies. The strengths of the manuscript are its public repository, the use of standard depth metrics, and its explicit discussion of limitations (e.g., the frame-selection module was not tested, and reconstruction could not be quantitatively aligned). However, the central claim of an effective integrated pipeline is directly contradicted by the paper's own Section 4.3, and the quantitative depth comparison is weakened by per-sequence scale fitting. As it stands, the significance is limited to a component-level comparison and a documentation of unsolved integration challenges; the claims need to be substantially revised or supplemented with quantitative reconstruction evidence.","major_comments":[{"comment":"The abstract's claim that 'Experiments conducted on the Hamlyn dataset demonstrate the effectiveness of the integrated methods' is directly contradicted by Section 4.3, which states that a quantitative comparison 'is not feasible at this point because the discrepancy between the reconstructed scene and the actual scene is too large,' and that the depth-estimation-based reconstruction in Figure 13 is 'far worse' than the ground-truth-depth reconstruction. The central claim of an effective integrated reconstruction pipeline is therefore unsupported by the paper's own results. The authors should either provide a quantitative reconstruction error measurement (e.g., point-to-surface distance after alignment) for the full pipeline, or explicitly downgrade the paper's claims to component-level depth estimation and ICP investigations.","section":"Abstract / Section 4.3"},{"comment":"The quantitative depth comparison is weakened by per-sequence scaling factors fitted retrospectively to align predicted depth maps with ground truth, as stated in Section 4.1 ('Scaling factors were applied to align the predicted depth maps with the ground truth') and Section 5 ('we must attempt a retrospective best fit scaling to align them with the ground truth'). With the scale fitted to the same data on which RMSE, MAE, and other metrics are computed, the reported numbers partly measure the quality of the fitted scale rather than an independent prediction. Please report scale-invariant metrics (e.g., SILog or median-scaling with held-out data), state exactly how many scale parameters are fitted per sequence, and discuss how the ranking changes under an alternative calibration.","section":"Section 4.1 / Section 5"},{"comment":"The 'best' ICP thresholding scheme is selected based on experiments on a single pair of frames from the test22 sequence using ground-truth depth images. This is insufficient to establish a generalizable recommendation and risks overfitting, especially because the same dataset and parameter choice are later used to support the qualitative reconstruction claims. Please evaluate the thresholding schemes on multiple frame pairs and sequences, and provide statistics (e.g., final alignment error, convergence trajectories) rather than a single anecdotal comparison.","section":"Section 4.2"},{"comment":"The manuscript explicitly states in Section 3 that the frame-selection module 'was not tested as part of the current pipeline evaluation.' Since frame selection is a named component of the integrated pipeline, the claim that the integrated methods are effective cannot be validated end-to-end. The authors should either incorporate and evaluate the frame-selection stage in the pipeline experiments, or clearly restrict the conclusions to the independent depth-estimation and ICP components.","section":"Section 3"}],"minor_comments":[{"comment":"The heading 'Quantative study of depth estimation methods' contains a typo; it should read 'Quantitative study of depth estimation methods.'","section":"Section 3 heading"},{"comment":"The tables report 'Mean' metric values but do not specify the number of frames over which the mean is computed or the standard deviation/variance of the metrics; adding these would allow the reader to assess the stability of the reported differences.","section":"Tables 1 and 2"},{"comment":"The SSIM data range is defined as max(max(dpred), max(dgt)), which is an unusual choice; please justify it or use a fixed, standard data range and state it explicitly.","section":"Equation (8)"},{"comment":"The caption describes the 25th frame of rectified01 as having 'erroneous ground truth' without a verification protocol; state how such frames are identified and whether they were excluded from the quantitative results in Tables 1 and 2.","section":"Figure 4"},{"comment":"The phrase 'the ground truth images are aligned okay' is informal and non-quantitative; if alignment quality is being claimed, it should be supported by a measured error (e.g., mean point-to-point distance or overlap ratio) rather than visual impression alone.","section":"Section 4.3"}],"recommendation":"major_revision","confidential_remarks":"This manuscript reads as a candid technical report rather than a completed journal article. Its main weakness is the internal contradiction between the abstract's 'effectiveness of the integrated methods' claim and Section 4.3's admission that reconstruction cannot be quantitatively aligned and is 'far worse' than ground-truth reconstruction. I recommend major revision rather than rejection because the authors are unusually transparent about limitations, they provide a public code repository, and a reframed component-level paper could be acceptable. The key risks to address are the per-sequence fitted scaling in the depth comparison, the single-frame-pair ICP threshold tuning, and the untested frame-selection module; the editors should require either new quantitative reconstruction results or a substantially narrowed set of claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: read Section 4.3 before the abstract. The body is transparent about what works and what doesn't, but the abstract's claim that the integrated pipeline is 'effective' does not survive contact with the results.\n\nWhat is actually useful: a modular pipeline with released code, a quantitative comparison of Depth-Anything V2, EndoDAC, and Depth-Anything V1 on two Hamlyn sequences, and a threshold sweep for ICP. The reproducibility details are good - config files, code paths, formulas, and an honest note that the frame-selection module was not tested. The depth comparison shows DA:v2 leading on this data, which is plausible and consistent with prior work.\n\nThe load-bearing problem is self-inflicted. Section 4.3 says a quantitative comparison of the reconstruction 'is not feasible at this point', that a single scaling cannot align the point clouds, and that the DA:v2 reconstruction is 'far worse' than one built from ground-truth depth. That directly contradicts the abstract's 'effectiveness of the integrated methods'. The depth tables themselves are also weakened by per-sequence scaling fitted to ground truth: reported RMSE and delta values after fitting a scale are not independent prediction accuracy. The ICP threshold experiment runs on a single frame pair, so the 'best' threshold is anecdotal. And because frame selection was not tested, the pipeline's first stage has no end-to-end validation.\n\nNone of this kills the report as a component-level description. As an integrated pipeline result, it does not hold. But the author is candid about the limitations in the discussion - the relative-depth problem, missing pose estimation, and inconsistent ground truth are all named.\n\nWho this is for: people building modular baselines for intraoperative 3D reconstruction, and readers who value an honest failure analysis more than a polished demo. It is not evidence that the full pipeline works.\n\nRecommendation: if this comes to a journal or workshop as a technical report, it deserves referee time after a major revision: rewrite the abstract to match Section 4.3, downgrade the reconstruction claim, add a held-out scaling procedure or per-frame uncertainty, and either test frame selection or drop it from the pipeline description. With the current abstract, I would not accept it as is, but I would not desk reject it either.","headline":"Honest component-level engineering report whose own reconstruction section contradicts the abstract's 'effectiveness' claim; useful as a baseline, not as evidence the integrated pipeline works.","tokens_in":10062,"tokens_out":3552,"would_cite":false,"duration_ms":33021,"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":"This report claims that a modular divide-and-conquer pipeline for endoscopic 3D reconstruction—frame selection, depth estimation, and ICP alignment—can serve as a practical baseline for image-guided surgery, with Depth-Anything V2…","keywords":["image-guided surgery","endoscopic 3D reconstruction","Depth-Anything V2","EndoDAC","Iterative Closest Point","frame selection","monocular depth estimation","Hamlyn dataset"],"falsifier":"Recompute the depth metrics on the same Hamlyn frames without any depth scaling and also with per-frame scaling or scaling fit only on valid ground-truth pixels; if Depth-Anything V2 no longer beats EndoDAC and Depth-Anything V1 on RMSE, MAE, and delta accuracy, the paper's central depth-ranking claim is not settled.","tokens_in":9137,"feed_emoji":"🏥","tokens_out":3632,"duration_ms":39267,"temperature":0.7,"pith_summary":"The report is trying to establish that a modular pipeline, which splits endoscopic 3D reconstruction into frame selection, depth estimation, and point-cloud alignment, can be a workable foundation for tracking surgical changes. Its central experimental claim is that Depth-Anything V2 produces better depth estimates than the endoscopic-specialized EndoDAC and the earlier Depth-Anything V1 on the Hamlyn dataset, with consistently lower RMSE, MAE, and related errors. The report also argues that adaptive thresholding in the Iterative Closest Point alignment—specifically using the mean plus two standard deviations of correspondence distances—yields the best alignment among the schemes tested. It openly acknowledges that monocular depth gives only relative scales, so end-to-end reconstruction remains qualitative rather than metric, and that pose estimation is the missing piece for clinical usefulness.","feed_headline":"Depth-Anything V2 beats EndoDAC on endoscopic depth in a modular pipeline","feed_subtitle":"DA:v2 gives lower Hamlyn depth errors in the chain, yet relative depth still blocks metric reconstruction.","key_machinery":"The central mechanism is a three-stage divide-and-conquer pipeline. Frame selection filters input frames using a HyperIQA quality score threshold and a red-channel intensity heuristic. Depth estimation then maps each selected frame to a depth map, with Depth-Anything V2 as the primary model and EndoDAC or Depth-Anything V1 as alternatives, followed by post-processing such as disparity-to-depth conversion and normalization. The 3D reconstruction stage converts depth maps to point clouds and aligns them with an Iterative Closest Point routine that uses SVD-based least-squares rigid alignment and a dynamic distance threshold, adaptively chosen as the mean plus two standard deviations of correspondence distances, to exclude outliers before computing the rotation and translation.","core_discovery":"The paper's central discovery, on its own terms, is that a general-purpose monocular depth model, Depth-Anything V2, provides better depth estimation results than both a state-of-the-art endoscopic depth model (EndoDAC) and its predecessor (Depth-Anything V1) on the Hamlyn test sequences. Across the reported tables, Depth-Anything V2 achieves lower Root Mean Squared Error, Mean Absolute Error, Squared Relative Error, and Log RMSE, and higher delta accuracy and SSIM, on both test22 and rectified01. The report further finds that among seven ICP thresholding schemes, the adaptive mean-plus-two-standard-deviations threshold gives the best alignment in qualitative reconstruction tests. It does not claim a metric-accurate 3D reconstruction; the final point clouds align visually with ground-truth input but not quantitatively, because relative monocular depth cannot be brought to an absolute scale by a single global scaling factor.","pith_inferences":["The report leaves implicit that the same adaptive statistical thresholding used in ICP could double as a frame-quality filter, flagging frames whose depth maps produce outlier-heavy correspondence distances; this is a testable extension that would reuse existing pipeline components.","A natural next step, which the report does not implement, is to couple the pipeline with a pose-estimation module; if pose were available, the neighbor-ICP drift problem would likely be replaced by a global optimization, converting the relative-depth outputs into a metric scene model.","The success of Depth-Anything V2 on two Hamlyn sequences suggests a broader hypothesis: that zero-shot general monocular depth models may transfer to endoscopic domains more effectively than domain-trained models when endoscopic training data are limited, but the report's two-sequence evidence is too thin to establish that generality.","The paper's concession that a single retrospective scaling factor cannot align relative depth to ground truth implies that per-frame or per-pixel scale estimation, or metric-depth training, is the true bottleneck; testing metric-depth models directly on Hamlyn would be a sharper comparison than the reported scaled-relative evaluation."],"forward_implications":["If the depth-ranking claim holds, a general monocular depth model can replace endoscopic-specialized depth networks in this pipeline, simplifying deployment and making future depth-model swaps a configuration change rather than a re-engineering effort.","The adaptive mean-plus-two-standard-deviations ICP threshold is identified as the best of seven schemes, giving future users a default setting for neighbor-based point-cloud alignment.","Because the end-to-end reconstruction is only qualitatively aligned and cannot be compared to ground truth at metric scale, the pipeline in its current form is a baseline, not a clinical tracking solution; adding pose estimation is a direct prerequisite for quantitative evaluation.","The frame selection module is implementable but untested on the curated Hamlyn sequences; its value is expected to appear on less curated operating-room footage with off-site frames, blur, or fluid interference.","The observed error spikes tied to erroneous ground-truth frames indicate that validation on endoscopic datasets should include a ground-truth quality-screening step before computing depth metrics."],"supporting_citations":[{"why":"Supplies Depth-Anything V2, the primary depth estimation model whose reported superiority over EndoDAC and Depth-Anything V1 is the paper's central quantitative claim.","marker":"[19]"},{"why":"Supplies EndoDAC, the state-of-the-art endoscopic depth model used as the main comparison baseline in the depth estimation experiments.","marker":"[21]"},{"why":"Supplies HyperIQA, the blind image quality assessment model used in the frame selection stage to score and threshold frames.","marker":"[18]"},{"why":"Supplies the Iterative Closest Point registration method that the reconstruction stage builds on for aligning point clouds.","marker":"[23]"},{"why":"Supplies the SVD-based least-squares rigid alignment used inside the custom ICP implementation to estimate rotation and translation from correspondences.","marker":"[22]"},{"why":"Supplies the standard evaluation protocol and metrics (RMSE, MAE, delta accuracy) used for the quantitative depth estimation comparison.","marker":"[27]"}],"fun_headline_variants":["Depth-Anything V2 beats EndoDAC on endoscopic depth","General depth model outperforms specialized endoscopic one","Depth-Anything V2 wins on Hamlyn depth for endoscopy","Modular pipeline: Depth-Anything V2 improves endoscopic depth"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The quantitative depth comparison assumes that a single retrospectively fitted scaling factor per sequence, applied after prediction, makes the monocular relative depth maps comparable to the Hamlyn ground truth; if that scaling absorbs systematic errors, the reported superiority of Depth-Anything V2 could be an artifact of the alignment procedure.","fun_headline_variants_meta":{"raw":{"variants":["Depth-Anything V2 beats EndoDAC on endoscopic depth","General depth model outperforms specialized endoscopic one","Depth-Anything V2 wins on Hamlyn depth for endoscopy","Modular pipeline: Depth-Anything V2 improves endoscopic depth"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000963,"raw_usage":{"total_tokens":4049,"prompt_tokens":842,"completion_tokens":3207,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":458,"completion_tokens_details":{"reasoning_tokens":3150}},"tokens_in":458,"tokens_out":3207,"duration_ms":22140,"temperature":1.0,"reasoning_tokens":3150,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:46:12.423995+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the depth metrics on the same Hamlyn frames without any depth scaling and also with per-frame scaling or scaling fit only on valid ground-truth pixels; if Depth-Anything V2 no longer beats EndoDAC and Depth-Anything V1 on RMSE, MAE, and delta accuracy, the paper's central depth-ranking claim is not settled.","supporting_citations":[{"cited_title":"Depth Anything V2,","cited_arxiv_id":null,"evidence_quote":"Supplies Depth-Anything V2, the primary depth estimation model whose reported superiority over EndoDAC and Depth-Anything V1 is the paper's central quantitative claim."},{"cited_title":"EndoDAC: Efficient Adapting Foundation Model for Self-Supervised Depth Estimation from Any Endoscopic Camera,","cited_arxiv_id":null,"evidence_quote":"Supplies EndoDAC, the state-of-the-art endoscopic depth model used as the main comparison baseline in the depth estimation experiments."},{"cited_title":"Blindly assess image quality in the wild guided by a self-adaptive hyper network,","cited_arxiv_id":null,"evidence_quote":"Supplies HyperIQA, the blind image quality assessment model used in the frame selection stage to score and threshold frames."},{"cited_title":"A method for registration of 3-d shapes,","cited_arxiv_id":null,"evidence_quote":"Supplies the Iterative Closest Point registration method that the reconstruction stage builds on for aligning point clouds."},{"cited_title":"Least-Squares Fitting of Two 3-D Point Sets,","cited_arxiv_id":null,"evidence_quote":"Supplies the SVD-based least-squares rigid alignment used inside the custom ICP implementation to estimate rotation and translation from correspondences."},{"cited_title":"Depth Map Prediction from a Single Image using a Multi-Scale Deep Network,","cited_arxiv_id":null,"evidence_quote":"Supplies the standard evaluation protocol and metrics (RMSE, MAE, delta accuracy) used for the quantitative depth estimation comparison."}],"review_version":1}