{"id":"e9dfccfb-0790-47b4-8366-ec1efdf47056","arxiv_id":"2506.24074","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"C3VDv2 releases 169 registered colonoscopy videos with depth, normals, optical flow, occlusion, pose, and 3D model ground truth, plus eight full-colon screening videos and fifteen deformation videos with enhanced realism.","lead":"A new dataset, C3VDv2, provides 192 colonoscopy videos with pixel-level 3D ground truth captured from realistic silicone colon phantoms. It adds fecal debris, blood, mucus, lens occlusion, and fast motion to previous clean-scene datasets, giving computer vision researchers a harder benchmark for 3D reconstruction in colonoscopy.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The unverified 2D/3D registration transform is load-bearing: all pixel-level ground truth inherits any alignment error, yet no independent geometric check exists on real videos and the validation metric reuses the optimized edge-overlap cue.","rationale":"The reader's weakest assumption correctly identifies the registration transform as the load-bearing element. The paper is transparent about the absence of a known true transform and about using edge overlap for both optimization and validation, so the concern is not a misreading. An independent geometric check, such as CT or fiducial-based pose verification, would either substantiate or refute the concern: if it passes, the conditional can be relaxed; if it fails, all pixel-level ground truth for that subset is compromised. I do not recommend rejection because the dataset is public, the protocol is clearly described, and the limitations are explicit; a conditional acceptance with a required verification experiment matches the evidence. This is the same adjustment the reader proposed, hence UNCHANGED.","tokens_in":16722,"tokens_out":5030,"duration_ms":63828,"concrete_test":"Acquire a CT scan of one C3VDv2 phantom segment mounted in its mold, segment the lumen surface, register the CT surface to the released 3D model to obtain an independent model-to-scan transform, and compare it with the release registration transform used for the corresponding clean-colon videos. If the two transforms differ by more than about 1 degree in rotation or 1 mm in translation, the edge-overlap objective is not a trustworthy proxy for true alignment and the pixel-level ground truth inherits a bias that the current Qalign metric cannot detect.","verdict_should_be":"UNCHANGED","load_bearing_attack":"C3VDv2's pixel-level ground truth is produced by rendering a virtual camera through the 3D model along a trajectory whose absolute position depends on one rigid transform T estimated by CMA-ES edge alignment. The paper states in Technical Validation that 'the true transformation for real videos is unknown.' The quantitative score Qalign is computed from Dice and Chamfer distances between the same DexiNed RGB edges and rendered depth edges that the registration optimizes; a high score mainly confirms convergence of the optimizer to a local maximum of its own objective, not correctness against physical truth. DexiNed edges include texture, specularity, and shadow boundaries absent from depth edges, so the objective can be systematically biased by appearance rather than geometry. Prior synthetic validation from C3VD19 used known transforms and clean rendered inputs, so it does not cover real phantom manufacturing defects, lighting, or debris. Qualitative overlay review is independent in modality but can miss a global rotation or scale error that appears plausible to a human. Consequently, every depth, normal, flow, occlusion, and pose map for a video inherits any error in T. This does not invalidate the dataset, but it makes 'pixel-level registered ground truth' conditional on an unverified calibration of the registration transform.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents C3VDv2, a colonoscopy 3D video dataset containing 192 videos (169,371 frames) acquired from silicone colon phantoms with a clinical colonoscope mounted on a robotic arm. For 169 videos, the authors provide per-frame ground truth depth, surface normals, optical flow, occlusion, diffuse maps, 6-DoF pose, and 3D coverage meshes, generated by registering a 3D phantom model to video frames via an edge-based 2D/3D alignment (CMA-ES optimization of an edge similarity score). The dataset includes clean and debris-filled paired videos, simulated fecal debris, mucus, blood, foam, water/debris on the lens, fast camera motion, and en-face views; 15 deformation videos and 8 simulated screening videos are provided without pixel-level GT. Technical validation consists of an SSIM-based synchronization check for the paired clean/debris videos and a combination of qualitative scoring and a quantitative edge-overlap score (Dice and Chamfer distances) for the 2D/3D registration.","tokens_in":16950,"tokens_out":5435,"duration_ms":60840,"significance":"If the registration accuracy holds, C3VDv2 is a substantial contribution: it is one of the largest public colonoscopy datasets with pixel-level 3D ground truth under realistic artifacts, and its paired clean/debris setup enables controlled evaluation of robustness. The manuscript is strong in its detailed fabrication and acquisition protocols, open-source code, and honest enumeration of limitations (e.g., debris GT does not model debris, deformation videos lack pixel-wise GT). The cross-modal SSIM validation for synchronization is a good use of independent data. However, the load-bearing assumption that the edge-based registration recovers the true rigid transform for real phantom videos is not independently verified. The quantitative validation metric is computed from the same edge-overlap cue that is optimized, so it mainly demonstrates convergence of the optimizer, not correctness against physical truth. The qualitative scoring provides a modality-independent check, but it cannot rule out moderate global errors. These issues are addressable and do not invalidate the dataset, but they need to be addressed before the central claim is fully supported.","major_comments":[{"comment":"The quantitative alignment score Qalign (Eq. 5) is computed from DexiNed edges on the RGB frames and Canny edges on the rendered depth frames. This is the same cue family that the registration objective in Eqs. (3)-(4) maximizes, so a high Qalign mainly indicates that the CMA-ES optimizer converged to a local maximum of the edge-overlap objective. The paper should provide an independent geometric verification on real phantom videos. A practical approach would be to perform a synthetic re-validation on the C3VDv2 phantoms: render depth and RGB images from known transforms, run the proposed registration pipeline, and report translational and rotational errors. Alternatively, adding external fiducials to a subset of phantoms and measuring their projected positions would provide a direct check. The current reliance on C3VD19's synthetic validation is insufficient because the v2 phantoms have new manufacturing defects and the acquisition conditions differ.","section":"Technical Validation, '2D/3D Registration Scoring'"},{"comment":"The paper does not report how many of the 169 registered videos received qualitative score 1, 2, or 3, nor the distribution of Qalign values. Since the usage notes explicitly recommend using only score 1 and 2 videos for tasks requiring high pixel-level accuracy, readers need the per-score counts to determine the effective size of the high-quality subset. Please include a summary table or histogram with the number of videos per qualitative score and the Qalign statistics for each group.","section":"Data Records / Usage Notes"},{"comment":"The registration pipeline uses an initial transform T_init that is 'manually estimated using the model-video overlay tool', followed by CMA-ES refinement. If the initialization is poor or if the search bounds (±0.1 rad, ±7.5 mm) are too tight, the optimizer can converge to a wrong local optimum. The paper should report the sensitivity of Qalign to the initialization and to the choice of the five keyframes, or at least state how often the search bounds had to be adjusted. This is important because a small number of keyframes (5) and manual substitution of low-edge frames could bias the estimated transform.","section":"Methods, '2D/3D Registration Pipeline'"}],"minor_comments":[{"comment":"The abstract states that 'Ground truth depth ... are provided for 169 colonoscopy videos', which is potentially misleading for the v3 debris videos because the depth, normal, and occlusion maps are those of the underlying clean 3D model and do not reflect the debris. Please qualify the sentence (e.g., 'with ground truth maps derived from the clean 3D model').","section":"Abstract"},{"comment":"In the row for C3VDv2 (Pixel-wise GT videos), the 'Debris' column is checked, but the pixel-wise GT does not model the debris surface. A footnote should clarify that the GT maps correspond to the clean colon geometry.","section":"Table 1"},{"comment":"Figure 7b would be more informative if the box plot included the number of videos in each qualitative-score category, and if the axes were fully described in the caption (e.g., what the boxes and whiskers represent).","section":"Figure 7"},{"comment":"The statement that 'the true transformation for real videos is unknown' appears only in the Technical Validation section. It should also be prominently stated in the Usage Notes or Data Records so that all users are aware of this limitation before downloading the dataset.","section":"Technical Validation"},{"comment":"The polynomial in Eq. (1) omits α1, and the text later says α1=0. This is standard for the Scaramuzza model, but an explicit note in the text would avoid confusion for readers not familiar with the model.","section":"Methods, Eq. (1)"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the scope of a dataset-oriented journal, and the dataset is likely to be a valuable community resource. The main concern is the circularity of the quantitative registration validation; adding an independent check (fiducial-based or synthetic re-rendering of the v2 phantoms) is essential. I would not reject the paper, as the limitations are honestly stated, but the central claim of 'pixel-level registered ground truth' requires stronger verification than the current Qalign metric provides. Also, the authors should be careful about the presentation of debris-video GT, since it is derived from the clean model and could be misused."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nC3VDv2 is a real addition, not a rehash. The v2 content—debris, blood, mucus, foam, water-on-lens, lens cleaning, water jet/suction, deformation clips, simulated screening videos, and paired clean/debris sequences—is exactly what the endoscopy 3D reconstruction field was missing. Prior phantom datasets were too clean; synthetic datasets lacked real optics and artifacts. The paired clean/debris design is especially nice: it lets you measure how much debris degrades a given method. The fabrication protocol is detailed enough to reproduce, and the data and code are public. The limitations section is unusually honest: v3 ground truth ignores debris, deformation videos have no pixel-wise GT, and score 3 videos are flagged.\n\nThe soft spot is the registration validation. The rigid transform T is load-bearing: every depth, normal, flow, occlusion, and pose map for a video is rendered through it. The paper estimates T by maximizing edge overlap (CMA-ES), and the quantitative Qalign score is computed from the same edge-overlap cue (DexiNed edges vs rendered depth edges). A high Qalign therefore mostly confirms that the optimizer converged to a local optimum of its own objective; it is not independent evidence that T matches physical truth. The paper says the true transform for real videos is unknown, and the prior synthetic validation from C3VD19 used known transforms and clean inputs, so it doesn't cover phantom defects, lighting, or debris. DexiNed edges include texture, specularity, and shadow boundaries that don't appear in depth edges, so systematic bias is plausible. The qualitative overlay check is independent in modality but could miss a global rotation or scale error that looks plausible to a human.\n\nThis doesn't kill the paper. The resource itself is valuable, and the limitations are stated clearly enough that a careful user can treat the ground truth as conditional on registration quality. But the validation section should be reframed as a convergence/quality check, not as independent geometric validation. An easy fix would be a CT scan of one phantom to independently verify T, or at least explicitly labeling Qalign as a quality indicator. The manual keyframe substitution is a minor issue; it introduces subjectivity but is unlikely to be load-bearing.\n\nI'd send this to peer review. It deserves a serious referee, and I'd want it in the literature for anyone doing depth estimation, SLAM, or NeRF-style reconstruction in colonoscopy.","headline":"A genuinely useful, well-documented dataset extension whose only real weakness is that the registration validation metric is partly circular; still deserves peer review.","tokens_in":17523,"tokens_out":2292,"would_cite":true,"duration_ms":24756,"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":"C3VDv2 expands colonoscopy benchmarking from clean clips to 192 videos with per-frame 3D ground truth and realistic clinical artifacts.","keywords":["colonoscopy video dataset","3D reconstruction ground truth","pixel-level depth and normals","optical flow","silicone colon phantom","2D/3D registration","endoscopic SLAM validation","simulated clinical artifacts"],"falsifier":"Take one of the phantom segments after imaging, scan it with CT or a structured-light scanner, and compare the scanned surface to the dataset's 3D model under the published per-video transform; if the residual misalignment on even the highest-quality videos exceeds a few millimeters, the registration is biased and the downstream pixel-level ground truth is in question.","tokens_in":16523,"feed_emoji":"🩺","tokens_out":7808,"duration_ms":82983,"temperature":0.7,"pith_summary":"C3VDv2 is a public video dataset built to give 3D-reconstruction researchers something existing colonoscopy datasets lack: per-pixel ground truth in videos that look and move like real clinical colonoscopy. The paper's claim is that 192 sequences, including 169 with pixel-level depth, normals, optical flow, occlusion, diffuse shading, 6-DoF poses, and 3D coverage models, capture artifacts such as fecal debris, blood, mucus, and lens occlusion, along with fast motion, loops, and paired clean-versus-debris trajectories. If the claim holds, colonoscopy SLAM, depth estimation, and reconstruction algorithms can be measured quantitatively under clinically relevant degradation rather than only on clean or synthetic scenes. The dataset also includes eight full-colon screening videos with recorded poses and fifteen deformation videos.","feed_headline":"C3VDv2 gives 169 colonoscopy videos pixel-level 3D ground truth","feed_subtitle":"Depth, normals, flow, and poses come registered to realistic phantoms, with debris, blood, and fast motion included.","key_machinery":"The load-bearing mechanism is a 2D/3D registration pipeline that fits a rigid transform $T$ aligning the known 3D colon model to video frames. The optimization minimizes $(1 - S_e)$, where $S_e$ is the normalized overlap between binary edges from RGB keyframes and edges rendered from depth maps of the 3D model, and the search is carried out with CMA-ES over rotations of $\\pm0.1$ radians and translations of $\\pm7.5$ millimeters. A spherical omnidirectional camera model maps each pixel to a ray, reproducing the colonoscope's wide field of view, and hand-eye calibration converts 500 Hz robotic-arm poses into camera poses. Because the phantom stays fixed and the robotic arm repeats its trajectory with sub-millimeter precision, the same rigid transform is reused for the paired debris-filled video, so pixel-level ground truth can be rendered even when debris obscures the scene.","core_discovery":"The central discovery is that high-fidelity silicone phantoms, robotic-arm trajectory logs, and edge-based 2D/3D registration can together produce pixel-level registered ground truth for colonoscopy videos that include realistic artifacts, at seven times the scale of the prior dataset. The paper demonstrates this on 169 registered videos totaling 67,886 frames, with per-frame depth, surface normals, optical flow, occlusion maps, diffuse maps, and camera poses, plus 3D coverage meshes. It further records eight simulated screening colonoscopies performed by an experienced gastroenterologist and fifteen deformation videos. The contribution is the dataset itself as a benchmark: 3D reconstruction and SLAM methods can now be quantitatively evaluated under fecal debris, blood, mucus, water and debris on the lens, fast and abrupt motion, and other conditions previously absent from paired ground-truth colonoscopy datasets.","pith_inferences":["Because the debris-filled videos inherit the clean colon's geometry as ground truth, an algorithm that reconstructs the debris surface itself will appear wrong; the intended reading is that the benchmark measures recovery of the underlying clean surface through visual clutter.","The loop and mirrored trajectories with known start-end correspondence make the dataset a ready-made drift benchmark: a SLAM system's endpoint pose error can be read directly from the provided poses.","The paired clean-versus-debris videos allow a controlled experiment of training on clean frames and testing on debris frames, isolating the effect of artifacts on depth and pose accuracy.","An independent geometric verification of the registration, such as CT-scanning one phantom after imaging and comparing the scanned surface to the published 3D model under the released transforms, would strengthen the dataset's accuracy claims, since the current quantitative validation uses the same edge-overlap cue that the registration optimizes."],"forward_implications":["Depth, normal, optical-flow, and occlusion maps give pixel-level supervision for monocular depth estimation, normal prediction, and flow networks in colonoscopy-like images.","The provided 6-DoF poses and coverage meshes turn every registered video into a quantitative test bed for endoscopic SLAM, NeRF, and 3D Gaussian Splatting reconstruction.","Paired clean and debris-filled videos recorded on the same trajectory make it possible to measure exactly how much debris degrades reconstruction and tracking performance.","Fast motion, loops, en-face-to-down-the-barrel transitions, and lens occlusion provide stress cases that clean colon datasets cannot provide.","Eight full-colon withdrawal videos with poses support validation of tracking on realistic screening trajectories, and fifteen deformation videos support qualitative robustness checks."],"supporting_citations":[{"why":"Provides the original C3VD dataset and the edge-based 2D/3D registration method that C3VDv2 reuses.","marker":"19"},{"why":"EndoSLAM is the main real-colonoscope comparison dataset, supplying real video and 3D models but not per-pixel registered ground truth.","marker":"14"},{"why":"EndoMapper supplies an in-vivo colonoscopy dataset and the calibration target used for photometric calibration.","marker":"15"},{"why":"DexiNed deep edge detection produces the binary edge maps that drive the registration loss.","marker":"53"},{"why":"The spherical omnidirectional camera model defines how pixels map to rays for rendering fisheye ground truth.","marker":"56"},{"why":"CMA-ES optimizes the rigid transform that aligns the 3D model to the video keyframes.","marker":"57"},{"why":"Hand-eye calibration converts robotic-arm poses into camera poses through solving AX=XB.","marker":"55"},{"why":"Inpainting removes specular reflections before edge extraction so they do not create spurious edges.","marker":"51"}],"fun_headline_variants":["169 colonoscopy videos with pixel-level 3D ground truth","C3VDv2: realistic phantom colonoscopy with dense 3D ground truth","Colonoscopy 3D dataset adds blood, debris, and fast motion to ground truth","Realistic phantoms bring 3D ground truth to 169 colonoscopy videos","C3VDv2: 169 videos with depth, normals, flow, and poses for colonoscopy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole dataset assumes the edge-based 2D/3D registration lands on the true phantom-to-camera transform for every video; since the true transform for real phantom videos is unknown and the validation uses the same edge-overlap signal that the optimization maximizes, a biased or locally stuck alignment would silently corrupt all pixel-level ground truth for that video.","fun_headline_variants_meta":{"raw":{"variants":["169 colonoscopy videos with pixel-level 3D ground truth","C3VDv2: realistic phantom colonoscopy with dense 3D ground truth","Colonoscopy 3D dataset adds blood, debris, and fast motion to ground truth","Realistic phantoms bring 3D ground truth to 169 colonoscopy videos","C3VDv2: 169 videos with depth, normals, flow, and poses for colonoscopy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000468,"raw_usage":{"total_tokens":2347,"prompt_tokens":973,"completion_tokens":1374,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":589,"completion_tokens_details":{"reasoning_tokens":1262}},"tokens_in":589,"tokens_out":1374,"duration_ms":13042,"temperature":1.0,"reasoning_tokens":1262,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:24:26.660556+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take one of the phantom segments after imaging, scan it with CT or a structured-light scanner, and compare the scanned surface to the dataset's 3D model under the published per-video transform; if the residual misalignment on even the highest-quality videos exceeds a few millimeters, the registration is biased and the downstream pixel-level ground truth is in question.","supporting_citations":[{"cited_title":"Bobrow, Gulfize Coskun, Kagan Incetan, Yasin Almalioglu, Faisal Mahmood, Eva Curto, Luis Perdigoto, Marina Oliveira, Hasan Sahin, Helder Araujo, Henrique Alexandrino, Nicholas J","cited_arxiv_id":null,"evidence_quote":"EndoSLAM is the main real-colonoscope comparison dataset, supplying real video and 3D models but not per-pixel registered ground truth."},{"cited_title":"Endomapper dataset of complete calibrated endoscopy procedures.Scientific Data, 10(1):671, 2023","cited_arxiv_id":null,"evidence_quote":"EndoMapper supplies an in-vivo colonoscopy dataset and the calibration target used for photometric calibration."},{"cited_title":"Soria, E","cited_arxiv_id":null,"evidence_quote":"DexiNed deep edge detection produces the binary edge maps that drive the registration loss."},{"cited_title":"Robot sensor calibration: solving ax= xb on the euclidean group.IEEE Transactions on Robotics and Automation, 10(5):717–721, 1994","cited_arxiv_id":null,"evidence_quote":"Hand-eye calibration converts robotic-arm poses into camera poses through solving AX=XB."},{"cited_title":"An image inpainting technique based on the fast marching method.Journal of graphics tools, 9(1):23–34, 2004","cited_arxiv_id":null,"evidence_quote":"Inpainting removes specular reflections before edge extraction so they do not create spurious edges."}],"review_version":1}