{"id":"f3d0a3c4-bb75-42d6-985b-fe60f7671f5e","arxiv_id":"2505.01838","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"MVHumanNet++ is a very large multi-view human dataset whose scale and annotations improve downstream 3D human reconstruction and generation models.","lead":"MVHumanNet++ is a large multi-view human video dataset with about 4,500 identities, 9,000 outfits, and over 645 million frames, adding normal and depth maps to the earlier MVHumanNet collection. The paper shows that increasing training data scale improves human reconstruction and generation models, pointing to the potential of large human data.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unvalidated mask post-processing can delete real body parts in the loose-clothing regime; annotation quality for the dataset's core use case is not established.","rationale":"The reader's conditional verdict is appropriate. I focused on the annotation pipeline because the paper's own text identifies the risk: Sapiens masks degrade on loose clothing, normal maps are pseudo-labels from Sapiens, and depth maps are rendered from 2DGS optimized with those normals. None of these is checked against ground truth. Algorithm 1 is the only mask-specific safeguard, and its largest-contour rule is mathematically guaranteed to drop any disconnected true body component. Since daily dressing and loose garments are the dataset's defining niche, errors here undermine the central claim more directly than the also-real inconsistencies in motion-sequence counts (60,000 versus 4,500 identities times two outfits times seven sequences) and release status (abstract says publicly available, while the conclusion says 'plan to release'). The scaling experiments at 100/2000/5000 outfits provide some evidence of utility, but they cannot validate annotation accuracy because training and testing share the same pseudo-label distribution; improvements could reflect fitting to the same systematic errors. A stratified manual-mask IoU audit with a depth reprojection check would settle the concern. If the audit shows high IoU and consistent depth, the central claim survives; if not, the dataset's annotation quality should be conditioned on a filtered subset or explicitly downgraded.","tokens_in":19519,"tokens_out":8468,"duration_ms":92331,"concrete_test":"Take a stratified sample of about 200 sequences from the 9,000 outfits, oversampling loose-fitting garments and self-occluding poses such as raised arms, wide skirts, and hands on hips. Manually correct 1,000 randomly selected frames using the manual labeling path already allowed in Algorithm 1, then compare Algorithm 1 output with the corrected mask using per-part IoU for arms, legs, and torso. If mean arm or leg IoU falls below 0.90, or if any frame loses more than 5% of its true foreground area to the largest-contour step, the mask pipeline is deleting body parts in the target regime. Additionally, render the 2DGS depth maps into neighboring cameras and measure reprojection error; systematic error above a few pixels would indicate the depth pseudo-labels are not reliable supervision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that MVHumanNet++ provides reliable annotations for 3D human digitization across 9,000 daily outfits rests on an unvalidated pseudo-label pipeline in Sec. 3.3. The weak point is Algorithm 1's Step 3: after extracting outer contours, it keeps only the largest contour and discards all others. In loose clothing, the exact regime the paper targets and where the authors admit Sapiens masks 'often exhibit noticeable artifacts', human silhouettes can split into multiple connected components, e.g., an arm separated from the torso by a segmentation error or a flared skirt occluding legs. Those components are deleted rather than merged or repaired, so the resulting masks can omit genuine body parts. Because masks feed the normal/depth processing and any mask-conditioned fitting, the error propagates. The paper provides only qualitative figures (Figs. 6, 8, 9) for mask, normal, and depth quality; there is no per-frame comparison against manual ground truth or a multi-view consistency check. Hence the utility of the entire annotation suite for the dataset's stated purpose is not established in the hardest, but central, loose-clothing cases.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces MVHumanNet++, a multi-view human capture dataset claimed to contain 4,500 human identities, 9,000 daily outfits, 60,000 motion sequences, and 645.1 million frames recorded with 48 or 24 synchronized RGB cameras. The dataset is advertised with annotations including human masks, camera parameters, 2D/3D keypoints, SMPL/SMPLX parameters, text descriptions, and newly added normal and depth maps. The authors describe the capture hardware, the recruitment and motion-protocol design, an automated annotation pipeline (Sapiens masks, OpenPose keypoints, EasyMocap SMPL fitting with VPoser, and 2DGS-rendered normal/depth maps), and a series of pilot studies on action recognition, generalizable NeRF and 3DGS reconstruction, text-driven image generation, generative models, and fine-tuning DUSt3R, reporting that larger training subsets yield improved downstream performance.","tokens_in":19731,"tokens_out":6099,"duration_ms":58528,"significance":"If the dataset is released with the claimed scale and usable annotation quality, it would be a valuable community resource for human-centric 3D vision, filling a gap between small-scale multi-view human datasets and large-scale object datasets. The paper's pilot studies provide a concrete demonstration of scaling benefits across several representative tasks, and the addition of normal and depth pseudo-labels is a useful extension over the original MVHumanNet. However, the significance rests on two unverified pillars: the exact scale and composition of the dataset (several counts are internally inconsistent) and the reliability of the automated annotations, especially for loose clothing, which is the regime the paper explicitly targets. The extensive experimental section is a strength, but the lack of quantitative annotation validation weakens the central claim.","major_comments":[{"comment":"The paper gives inconsistent counts of motion sequences. Section 1 states each participant is recorded in 'seven different motion sequences', while Section 3.2 describes a protocol of 12 action sets per participant (six in the first outfit and six in the second) plus in-place A-pose rotations. Neither the abstract's 60,000 motion sequences nor Table 1's 'Actions' column of 500 can be derived from these numbers (4,500 × 7 = 31,500; 4,500 × 12 = 54,000; with rotations the total would differ again). Please define what constitutes a 'motion sequence' (e.g., one performance of one action on one side, or a full outfit session) and reconcile all reported counts in the text, abstract, and Table 1.","section":"§1 and §3.2"},{"comment":"Algorithm 1's mask-enhancement step discards all outer contours except the largest one. In loose clothing, the exact regime where the authors note Sapiens masks 'often exhibit noticeable artifacts', human silhouettes can split into multiple connected components (e.g., an arm separated by a segmentation error or a flared skirt occluding the legs). The procedure would delete those components rather than merge them, possibly erasing genuine body parts. Because these masks feed the normal/depth generation (via 2DGS), any mask-conditioned fitting, and the released annotations, this post-processing choice is load-bearing. The paper provides only qualitative visualization (Figs. 6, 8, 9) and no quantitative evaluation (e.g., IoU against manual masks on a sample stratified by garment type, or a multi-view consistency check). Without such validation, the claim that MVHumanNet++ provides reliable masks for 3D human digitization across 9,000 daily outfits is not established.","section":"§3.3, Algorithm 1"},{"comment":"The depth maps are rendered from 2D Gaussian Splatting (2DGS) that was itself refined using Sapiens normal maps, and these same depth maps are then used as pseudo-labels to supervise EVA-Gaussian's depth estimator. The paper does not evaluate the accuracy or consistency of the depth pseudo-labels directly, for example by checking multi-view depth agreement or comparing rendered depth against a small set of manual/geometric ground-truth depth on controlled captures. Since the depth maps are presented as a new annotation modality that is 'a reliable source of geometric information', the absence of any direct geometric validation makes it impossible to assess their contribution independently of the downstream rendering metrics.","section":"§3.3, Depth Maps and §4.4"},{"comment":"The '#Actions' column in Table 1 is ambiguous and the comparison with DNA-Rendering is unverifiable. The MVHumanNet++ text says the dataset contains 500 motion types, while the abstract reports 60,000 motion sequences; Table 1 lists '500' in the #Actions column for MVHumanNet++ and '1187' for DNA-Rendering, yet the introduction claims MVHumanNet++ surpasses DNA-Rendering 'by an order of magnitude in terms of motion'. If #Actions counts action types, 500 < 1187; if it counts sequences, the column mixes units across datasets. Please specify the metric and use consistent units in all comparative statements.","section":"Table 1 and Introduction"},{"comment":"The abstract and introduction state that MVHumanNet++ 'is publicly available' at a URL, while the Conclusion says 'We plan to release the MVHumanNet++ dataset with annotations publicly'. These two statements directly contradict each other regarding the dataset's availability status. Please correct this inconsistency and clarify whether the dataset is currently downloadable and under what terms.","section":"Abstract and Conclusion"}],"minor_comments":[{"comment":"The manuscript contains several typos and ungrammatical phrases, including 'the emergency of' (should be 'the emergence of'), 'comparsion' in the caption of Fig. 7, and 'we refer human prior' in Section 4.2. A careful proofreading pass is needed.","section":"Throughout"},{"comment":"References [25] and [87] are the same paper (Pavlakos et al., 'Expressive body capture') and should be merged or cross-referenced. Also, reference [24] contains a typo 'T oG' instead of 'TOG'.","section":"References"},{"comment":"The text says experiments used '62% of the available data' and '2800 identities ... amounting to a total of 5500 sets', but 2800 identities × 2 outfits would give 5600 sets. The relation between 2800 identities, 5500 sets, and the 62% figure should be clarified.","section":"§4, first paragraph"},{"comment":"The action-recognition experiment reports training on 'approximately 4000 outfits' and testing on '400 outfits', whereas other experiments in Section 4 use 5500 outfits. Please clarify whether these are intentional subsets and how they relate to the overall 62% split.","section":"§4.1"}],"recommendation":"major_revision","confidential_remarks":"The internal contradictions in the reported dataset counts (7 vs 12 sequences per participant; 60,000 vs 31,500/54,000 total sequences) and the release-status discrepancy between the abstract and conclusion are unusual for a dataset paper and suggest the manuscript may need a thorough consistency pass before it can be evaluated for publication. The core annotation-quality concern (Algorithm 1's largest-contour-only rule) is technical and fixable in principle, but it requires additional experiments and possibly pipeline changes. I recommend major revision, not rejection, because the dataset could be a significant contribution if the counts are corrected and the annotation pipeline is validated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague--\n\nThe paper to know: MVHumanNet++ is a genuine but incremental extension of the authors' CVPR 2024 MVHumanNet. The headline numbers are the same (4,500 identities, 9,000 outfits, 60,000 sequences, 645M frames), and the '+' additions are improved masks and SMPL/SMPLX, pseudo-labeled normal and depth maps, plus several pilot studies. If the numbers hold up, this is the largest multi-view human capture dataset by a wide margin, and that scale is the main value.\n\nCredit where due: the authors are transparent about their annotation pipeline and give a concrete GPU-hour accounting. Algorithm 1 is explicitly specified, which means the community can inspect it and spot problems—that is how good dataset papers behave. The pilot studies (generalizable NeRF/3DGS, generative models, DUSt3R fine-tuning) are the right experiments to run, and the reported scaling trends are plausible even if not surprising. Adding VPoser to EasyMocap's SMPL fitting is a sensible fix to a known failure mode.\n\nThe soft spots are real. The reader's stress-test concern lands: Algorithm 1 discards all contours except the largest. For loose clothing, which the paper itself says is where Sapiens masks show artifacts, silhouettes can split into disconnected regions, so this step can remove an arm or a leg separated by a flared skirt. Because masks drive normal/depth generation and mask-conditioned fitting, the error propagates. The paper gives only qualitative figures for mask, normal, and depth quality—no per-frame comparison against manual ground truth, no multi-view consistency check. For a dataset whose central promise is reliable annotation for 3D digitization, this is the biggest gap.\n\nThere are mechanical inconsistencies too: the abstract says 'seven different motion sequences' per participant while Sec. 3.2 describes 12 action sets; the conclusion says 'plan to release' while the abstract and project page say publicly available; and the GPU-hour numbers are large but never reconciled with the frame count. All fixable, but a referee will want them cleaned.\n\nThe depth/normal circularity is mild: Sapiens normals refine 2DGS, whose rendered depth supervises EVA-Gaussian. That is a known bias, not a fatal flaw. The central scale claim is not circular; it is just under-validated.\n\nBottom line: this paper deserves a serious referee. My own verdict is skeptical about annotation quality in the loose-clothing regime, but positive about the dataset's potential value. I would bring it to our reading group and probably cite it as the large-scale human capture reference once the data is actually released and the mask pipeline is validated on a sample against manual ground truth.","headline":"MVHumanNet++ is a modest but real extension of the authors' CVPR 2024 dataset; the scale is valuable, but pseudo-label quality in loose clothing and internal inconsistencies need work before I'd trust the annotations fully.","tokens_in":20283,"tokens_out":3114,"would_cite":true,"duration_ms":30541,"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":"MVHumanNet++ claims the largest multi-view human capture dataset to date, with 4,500 identities, 9,000 outfits, and 645.1 million annotated frames, and shows that downstream reconstruction and generation models improve as training data…","keywords":["multi-view dataset","3D human digitization","human capture","SMPL/SMPLX annotation","neural rendering","3D generation","dataset annotation","normal and depth maps"],"falsifier":"Take a random sample of sequences with loose-fitting clothing (wide skirts, open jackets, loose trousers), hand-correct the masks on a few hundred frames, and measure how often the largest-contour step erases a detached limb, a flapping hem, or a separate garment piece; also compare the Gaussian-splatting-rendered depth maps against multi-view stereo or structured-light depth on a subset. If the error rate is high or the depth errors are systematic, the claimed utility of the annotations for reconstruction across all 9,000 outfits fails.","tokens_in":1906,"feed_emoji":"📹","tokens_out":2600,"duration_ms":94020,"temperature":0.7,"pith_summary":"MVHumanNet++ sets out to give the 3D human community the kind of large-scale resource that powered large language models and text-to-image models: 4,500 people in everyday clothing, captured by synchronized multi-view camera rigs, yielding 9,000 outfits, 60,000 motion sequences, and 645.1 million frames. Alongside raw images it provides masks, camera parameters, 2D and 3D keypoints, SMPL/SMPLX body-model parameters, action labels, textual descriptions, and newly added normal and depth maps. The paper's central claim is that this scale is what has been missing: pilot experiments show action-recognition accuracy, generalizable neural rendering and Gaussian-splatting reconstruction, text-driven human image generation, and 3D generative models all improve as more of the dataset is used for training. A sympathetic reader would take the contribution as evidence that human-centric 3D tasks can move from per-subject optimization to data-driven generalization, provided the automatic annotation pipeline holds up across daily clothing.","feed_headline":"4,500 identities, 645 million frames for 3D humans","feed_subtitle":"Daily outfits, 60,000 motions, and rich annotations aim to make human digitization scale like image and text AI.","key_machinery":"The central object is the MVHumanNet++ dataset itself: a synchronized multi-view capture arrangement (48 cameras at up to 12MP in the primary rig and 24 cameras at 5MP in the secondary rig) plus a largely automatic annotation pipeline. The pipeline runs in stages: human masks are produced by a human-vision segmentation model and refined by a contour post-processing algorithm, 2D keypoints and multi-view triangulation give 3D skeletons, a multi-view body-model fitting procedure is regularized by a variational body-pose prior to prevent twisted joints, and normal estimates feed a Gaussian-splatting surface optimization whose rendered depth becomes the dataset's depth maps. The work of this machinery is to turn raw multi-view video into a reusable, ready-to-train asset: the claimed value is that any downstream model improves predictably as more of these annotated outfits are fed in.","core_discovery":"The paper's central discovery is that a deliberately ordinary-data capture strategy, everyday outfits, 500 action types, two outfits per person, and short motion sequences in a calibrated 48- or 24-camera dome, can scale human performance capture to sizes an order of magnitude beyond the closest prior multi-view human dataset, and that the resulting annotations (masks, cameras, keypoints, SMPL/SMPLX parameters, text, normal maps, depth maps) convert that raw scale into measurable gains across 2D and 3D human tasks. Concretely, it reports that generalizable human reconstruction models trained on 5,000 outfits outperform the same models trained on 100 or 2,000 outfits, that pretraining on the dataset and fine-tuning on a smaller public benchmark beats training on that benchmark from scratch, that regenerated SMPL/SMPLX fits with a body-pose prior remove twisting artifacts and improve per-subject animatable reconstruction, and that generative models for 2D, 3D, and multi-view humans improve in image quality scores as training subjects increase. It frames the dataset as the current largest-scale 3D human dataset.","pith_inferences":["Editorial inference: if the mask and pose annotations are accurate on loose clothing, the dataset's 60,000 labeled motion sequences could support video-level generative models of humans, not just single-frame reconstruction, a use the paper does not demonstrate.","Editorial inference: because the depth maps are rendered from the same Gaussian-splatting fits that reconstruction models learn from, downstream gains on depth-supervised tasks may partly reflect alignment between training labels and model family rather than ground-truth geometry; a natural stress test is evaluating on independently captured depth.","Editorial inference: the paper's pattern of performance improving with scale on every tested task suggests the dataset may behave like a scaling-law resource for human digitization, but the experiments cover only 62% of the data, so whether the trend continues on the full 4,500 identities is untested.","Editorial inference: the manual text descriptions paired with multi-view images could be used for text-driven avatar generation and retrieval, an application the paper leaves mostly implicit beyond image generation."],"forward_implications":["Scaling training data from 100 to 5,000 outfits raises novel-view synthesis quality for generalizable neural rendering and Gaussian-splatting reconstruction methods, with image-quality scores improving for both scene-level and human-prior baselines.","A model pretrained on MVHumanNet++ and then fine-tuned on a smaller public human benchmark outperforms the same model trained on that benchmark from scratch.","The improved SMPL/SMPLX fits, regularized by a body-pose prior, yield higher-fidelity per-subject animatable Gaussian reconstruction than the earlier annotation version.","Text-conditioned, pose-conditioned human image generation and multi-view human generative models both improve when trained on the larger split of the dataset rather than the smaller split.","The dataset's rendered depth maps can serve as supervision for generalizable human reconstruction, reducing depth ambiguity when reconstructing humans from unconstrained image pairs."],"supporting_citations":[{"why":"Defines the prior conference version of the dataset and the baseline annotation approach that this work extends.","marker":"[29]"},{"why":"The concurrent multi-view human dataset that MVHumanNet++ compares against and claims to exceed by an order of magnitude.","marker":"[15]"},{"why":"Supplies the human-mask and normal-map seed estimates that feed the annotation pipeline.","marker":"[84]"},{"why":"Provides the 2D keypoints used for multi-view 3D skeleton triangulation.","marker":"[85]"},{"why":"Provides the multi-view SMPL/SMPLX fitting base that the authors regularize with a pose prior.","marker":"[86]"},{"why":"Supplies the variational body-pose prior that removes joint twisting and self-intersection artifacts in fitted bodies.","marker":"[87]"},{"why":"Renders the normal-refined surfaces from which the dataset's depth maps are derived.","marker":"[90]"},{"why":"The per-subject Gaussian-splatting method used to evaluate the improved SMPL/SMPLX annotations.","marker":"[55]"},{"why":"The uncalibrated two-view reconstruction model fine-tuned on MVHumanNet++ in the final pilot study.","marker":"[28]"}],"fun_headline_variants":["Largest 3D human dataset: 4,500 people, 645M frames","Daily outfits, 645M frames: scaling 3D human digitization","MVHumanNet++: ordinary clothes, 60K motions, 9K outfits","Record 3D human data: 4,500 identities, rich annotations"],"cache_read_input_tokens":22528,"weakest_assumption_plain":"The load-bearing premise is that the fully automatic annotation pipeline, automatically segmented human masks, triangulated keypoints, body-model fitting with a pose prior, and rendered normal and depth maps, produces labels accurate enough for every one of the 9,000 outfits, including loose clothing, without per-frame ground-truth checks; in particular, the mask post-processing step that keeps only the largest outer contour can silently erase body parts that appear detached in loose garments.","fun_headline_variants_meta":{"raw":{"variants":["Largest 3D human dataset: 4,500 people, 645M frames","Daily outfits, 645M frames: scaling 3D human digitization","MVHumanNet++: ordinary clothes, 60K motions, 9K outfits","Record 3D human data: 4,500 identities, rich annotations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000339,"raw_usage":{"total_tokens":1957,"prompt_tokens":1113,"completion_tokens":844,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":729,"completion_tokens_details":{"reasoning_tokens":756}},"tokens_in":729,"tokens_out":844,"duration_ms":7647,"temperature":1.0,"reasoning_tokens":756,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:09:15.360100+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a random sample of sequences with loose-fitting clothing (wide skirts, open jackets, loose trousers), hand-correct the masks on a few hundred frames, and measure how often the largest-contour step erases a detached limb, a flapping hem, or a separate garment piece; also compare the Gaussian-splatting-rendered depth maps against multi-view stereo or structured-light depth on a subset. If the error rate is high or the depth errors are systematic, the claimed utility of the annotations for reconstruction across all 9,000 outfits fails.","supporting_citations":[{"cited_title":"Mvhumannet: A large-scale dataset of multi-view daily dressing human captures,","cited_arxiv_id":null,"evidence_quote":"Defines the prior conference version of the dataset and the baseline annotation approach that this work extends."},{"cited_title":"Sapiens: Foundation for human vision models,","cited_arxiv_id":null,"evidence_quote":"Supplies the human-mask and normal-map seed estimates that feed the annotation pipeline."},{"cited_title":"Realtime multi-person 2d pose estimation using part affinity fields,","cited_arxiv_id":null,"evidence_quote":"Provides the 2D keypoints used for multi-view 3D skeleton triangulation."},{"cited_title":"Easymocap - make human motion capture easier","cited_arxiv_id":null,"evidence_quote":"Provides the multi-view SMPL/SMPLX fitting base that the authors regularize with a pose prior."},{"cited_title":"Expressive body capture: 3d hands, face, and body from a single image,","cited_arxiv_id":null,"evidence_quote":"Supplies the variational body-pose prior that removes joint twisting and self-intersection artifacts in fitted bodies."},{"cited_title":"2d gaussian splatting for geometrically accurate radiance fields,","cited_arxiv_id":null,"evidence_quote":"Renders the normal-refined surfaces from which the dataset's depth maps are derived."},{"cited_title":"Animatable gaussians: Learning pose-dependent gaussian maps for high-fidelity human avatar modeling,","cited_arxiv_id":null,"evidence_quote":"The per-subject Gaussian-splatting method used to evaluate the improved SMPL/SMPLX annotations."}],"review_version":1}