{"id":"f63f42db-4e0a-4142-82cc-13eb1b0df70b","arxiv_id":"2509.06607","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"SKEL rigs the SMPL body surface to a new 46-degree-of-freedom biomechanical skeleton, trained on BioAMASS, a new dataset of skeletons fitted inside SMPL meshes from AMASS.","lead":"The authors build SKEL, a 3D human body model that combines SMPL's skin with a biomechanically realistic skeleton, so one set of parameters drives both the body surface and the bones. They also release BioAMASS, a dataset of biomechanical skeletons fitted inside thousands of posed SMPL bodies, and show how SKEL can upgrade existing human pose datasets for biomechanics.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The accuracy claim depends entirely on the AddBiomechanics pseudo-ground-truth (Secs. 4.3, 6.2, 7); without external anatomical validation, 'more biomechanically accurate joint locations' is not actually established.","rationale":"The paper is well-executed and honest: it releases model, code, and data, acknowledges the pseudo-ground-truth limitation in Sec. 7, and reports concrete fit errors. The re-rigging construction (Sec. 5) is a useful contribution independent of the accuracy of the labels, and the claim of a single (beta, q) parameterization for skin and skeleton is supported by Eqs. (5)-(9). However, the headline claim that goes beyond engineering, namely 'biomechanically accurate joint locations', is only as strong as the labels. The reader's weakest_assumption identified exactly this issue. Reading Eq. (2) adds a concrete reason for concern: delta is a free per-marker offset in the same objective that determines bone scales and joint locations, so low marker residuals do not by themselves guarantee accurate joint centers. The evaluation in Sec. 6.2 uses the same AddBiomechanics fits as reference, making the accuracy assessment circular with respect to the pseudo-GT assumption. The paper's own wording in Sec. 7 and the acknowledged humerus error (Fig. 10) confirm the gap. Therefore the appropriate verdict remains CONDITIONAL: the model and dataset are valuable, but the 'biomechanical accuracy' claim needs independent validation or a scoped phrasing before being accepted at face value. I do not see an internally inconsistent argument that would justify REJECT; the limitation is an external validity gap rather than a soundness failure.","tokens_in":18368,"tokens_out":8108,"duration_ms":76358,"concrete_test":"Select a small cohort (e.g., 5-10 subjects) for whom both SMPL-fittable body meshes and subject-specific bone geometry from CT/MRI are available; generate the Sec. 4.3 virtual markers on their SMPL meshes, run AddBiomechanics to obtain BSM joint centers, and compare femur-head, humerus-head, tibia, and knee joint centers to imaging-derived anatomical landmarks. If the median Euclidean error approaches or exceeds the SMPL-vs-BSM difference reported in Fig. 11, or if humerus-head error exceeds about 20 mm, the pseudo-GT labels are biased and the 'biomechanically accurate' claim should be scoped down to 'consistent with marker-based optimization' rather than anatomical accuracy.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that SKEL yields 'more biomechanically accurate joint locations' than SMPL rests on one assumption: that the BSM fits produced by AddBiomechanics from virtual skin markers recover true anatomical joint centers. Everything downstream inherits that assumption. BioAMASS labels are created by solving Eq. (2) with a per-marker offset delta (Sec. 4.3.2); the joint regressor is trained on those labels (Sec. 5.1), and the quantitative comparisons in Sec. 6.2 and Figs. 10-11 evaluate against the same labels ('J^B that we consider ground truth'). If the marker-based optimization is biased, for example at the shoulder/humerus head or because delta absorbs systematic skin-to-bone offsets, the reported sub-centimeter regression errors measure consistency with a self-generated fit, not biomechanical accuracy. The paper itself flags this in Sec. 7: the fits 'should not be considered as actual ground truth, but rather a pseudo-ground truth.' No independent validation against medical imaging, bone-pin data, or functional joint-center measurements is provided. The qualitative OSSO comparison (Sec. 6.4) also does not quantify the claim that bones fit inside the body surface better. This gap is load-bearing because the central novelty is the biomechanical validity of the skeleton, not just the re-rigging machinery.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces SKEL, a parametric body model that re-rigs the SMPL surface model with a biomechanical skeleton, BSM, so that skin and skeleton are controlled by the same pose vector q and shape vector β. To train SKEL, the authors construct BioAMASS, a dataset of about 9 hours of motion from 113 AMASS subjects, by placing virtual markers on SMPL meshes and fitting BSM with the AddBiomechanics optimization pipeline. They learn a non-negative joint regressor from SMPL vertices to BSM anatomical joint locations, estimate per-bone base rotations and shape-dependent corrective rotations, and define a new kinematic tree with 46 biomechanical degrees of freedom, including a constant-curvature spine, sliding scapulae, and radioulnar pronation/supination. The paper evaluates joint-location errors against the BSM fits used as pseudo-ground-truth, reports SKEL-to-SMPL surface fitting errors on DFAUST, and shows qualitative comparisons with OSSO on MOYO.","tokens_in":18723,"tokens_out":5262,"duration_ms":50952,"significance":"If the accuracy claims are taken at face value, SKEL is a valuable bridge between statistical body models and biomechanics: it is, to my knowledge, the first model in which skin and a biomechanical skeleton are driven by the same pose and shape parameters, and the BioAMASS dataset and the SMPL-to-BSM fitting pipeline are potentially useful resources. The re-rigging formulation is well motivated, and the paper includes several thoughtful design choices, such as using OSSO to personalize virtual marker offsets and modeling the shoulder blade with an ellipsoid constraint. The main limitation is that the central claim of biomechanical accuracy is evaluated against the same AddBiomechanics fitting pipeline that generated the training labels, and the paper's own Sec. 7 acknowledges this is pseudo-ground-truth. Consequently, the sub-centimeter reported errors demonstrate reproducibility of a fitting pipeline rather than independently established anatomical accuracy; the qualitative OSSO comparison also does not quantify the 'bones fit inside the body surface better' claim.","major_comments":[{"comment":"The evaluation of the joint regressor and the SKEL fits is circular. The joint locations J^B used as 'ground truth' in Fig. 10 and Fig. 11 are produced by the same AddBiomechanics optimization (Eq. (2)) that generated the BioAMASS training labels. The learned regressor and SKEL are therefore scored by their ability to reproduce the fitting pipeline, not by their anatomical correctness. The paper itself states in Sec. 7 that these fits 'should not be considered as actual ground truth, but rather a pseudo-ground truth.' To support the abstract's claim of 'more biomechanically accurate joint locations than SMPL,' the authors should add an external validation, for example comparing a subset of subjects against OSSO skeletons trained on medical images, functional joint-center estimates, or published anthropometric or imaging measurements; alternatively, the claims throughout the abstract, Sec. 1, and Sec. 6.2 should be softened to 'consistency with a marker-based biomechanical fitting pipeline.'","section":"Sec. 6.2, Sec. 4.3.2, Sec. 7"},{"comment":"The claim that the bones 'fit inside the body surface better than previous methods' is supported only by qualitative side-by-side images of OSSO and SKEL on MOYO. No quantitative metric is reported for bone containment, bone-skin penetration, or agreement of bone orientations with an independent reference. I recommend adding a scalar evaluation, such as the fraction of bone vertices inside the SMPL skin mesh, the signed distance of bone vertices to the skin surface, or a comparison of joint angles against the AddBiomechanics reference, on a held-out set of poses and subjects.","section":"Sec. 6.4"},{"comment":"The per-marker offsets δ and the marker weights λ_k are free parameters in the bi-level optimization that creates BioAMASS, and the paper does not state how δ is initialized or regularized. Unrestrained per-marker offsets can absorb systematic skin-to-bone offsets and shift the estimated joint centers, so the skeleton locations might be determined more by δ than by the marker trajectories. Please report the magnitude and variability of the learned offsets, and provide a sensitivity analysis showing that joint locations in BioAMASS are stable when δ is removed or differently regularized.","section":"Sec. 4.3.2, Eq. (2)"}],"minor_comments":[{"comment":"The formula for t_spine is difficult to parse; the term '(r−cos(α)) ∗ −sin(q_z)' appears to have a missing operand or an unusual double sign, and the definitions of q_x, q_y, q_z are incomplete. Adding explicit parentheses and a short derivation would improve readability.","section":"Sec. 4.2, Eq. (1)"},{"comment":"The text refers to 'the green term' in the equation, but the printed equation is monochrome; please replace the color reference with an explicit description of the factors being discussed.","section":"Sec. 5.2, Eq. (6)"},{"comment":"Sec. 5.2 states that SKEL can match SMPL meshes 'with an average vertex-to-vertex error below 3 cm,' while Sec. 6.3 reports average mean differences of 1.1 cm for males and 0.9 cm for females on DFAUST. Please reconcile these numbers, and clarify whether the 3 cm figure refers to extreme poses outside DFAUST.","section":"Sec. 5.2 and Sec. 6.3"},{"comment":"The prior P(s,β) is said to use biological sex as in Werling et al., but the paper does not explain how biological sex is obtained from the SMPL shape parameters β. Please specify the sex estimation procedure or state that sex is assumed from the source dataset.","section":"Sec. 4.3.2"},{"comment":"The color bar is labeled 'Blue: 0 cm, Red: 2 cm,' yet the text reports a maximum female difference of 1.9 cm; please adjust the visualization range or the text so the figure and caption are consistent.","section":"Fig. 12"}],"recommendation":"major_revision","confidential_remarks":"The main risk is the gap between the headline claims of biomechanical accuracy and the self-referential evaluation. The technical construction of BioAMASS and SKEL is a solid contribution, and the authors are transparent about the pseudo-ground-truth limitation in Sec. 7. If the authors either add one external anatomical validation (e.g., a small comparison against OSSO or imaging-derived joint centers) or systematically soften the accuracy claims in the abstract and Sec. 1, the paper would be publishable. This is why I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nShort version: this is a real contribution. The self-referential evaluation concern is legitimate, but the authors are unusually honest about it. I'd send it to review.\n\nWhat's new: SKEL is the first parametric body model I know of where skin and an anatomical skeleton are driven by the same pose parameters, while inheriting SMPL's shape space. The BioAMASS dataset, built by fitting their new BSM skeleton to AMASS sequences via AddBiomechanics, is a useful resource on its own. The re-rigging machinery is clever and well described: the learned base bone rotations, the shape-dependent corrections, the constant-curvature spine, and the scapula sliding. The paper is also transparent about its limits — Section 7 explicitly calls the BSM fits \"pseudo-ground truth,\" flags the humerus head as a known weak spot, and says clinical validation is the necessary next step. That is the right kind of candor.\n\nWhere it's soft: the headline claim — \"more biomechanically accurate joint locations than SMPL\" — is measured against the same AddBiomechanics pipeline that produced the training labels. That is consistency with a model, not anatomical accuracy. The abstract and intro lean on \"biomechanically accurate\" fairly hard, even though the body of the paper scopes it properly. The OSSO comparison in Section 6.4 is qualitative only; a quantitative inside-the-skin test would strengthen it. And the pose-corrective transfer from SMPL is imperfect, which they acknowledge, but it means SKEL's surface deviates from SMPL by about a centimeter on average — so \"upgrading\" SMPL datasets to SKEL carries a small surface cost.\n\nThese are not fatal flaws. The contribution is the pipeline, the dataset, and the model, and those stand. The accuracy claim needs to be scoped as \"consistent with a state-of-the-art marker-based biomechanical fit,\" not \"validated against true bone positions.\" If the authors rephrase that and add one independent check — even a small medical-imaging or functional-joint-center dataset — the paper becomes much stronger.\n\nWho it's for: graphics and vision researchers who want biomechanical parameters out of SMPL-based systems, and biomechanists who want a shape space on top of OpenSim-style skeletons. Both communities will use it. I'd cite it as the current state of the art for skin-skeleton parametric models, with the caveat about validation.\n\nRecommendation: serious referee time. Conditional accept, with the accuracy claim tightened and ideally a small external validation.","headline":"A genuinely useful skin-skeleton parametric model with an accuracy claim that is self-referential but explicitly acknowledged; deserves serious peer review and a conditional accept.","tokens_in":19252,"tokens_out":1846,"would_cite":true,"duration_ms":17309,"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":"SKEL gives the standard SMPL human body model a biomechanical skeleton driven by the same shape and pose parameters as the skin surface.","keywords":["parametric human body model","biomechanical skeleton","SMPL","motion capture","joint regression","dataset generation","digital human animation","musculoskeletal model"],"falsifier":"Take a subject with both a CT or MRI scan and synchronized marker motion capture, fit SKEL to the motion, and compare its femur-head and humerus-head joint centers with the imaging-derived centers; a systematic offset, especially at the shoulder, would falsify the claimed biomechanical accuracy.","tokens_in":18174,"feed_emoji":"🦴","tokens_out":10060,"duration_ms":83373,"temperature":0.7,"pith_summary":"This paper tries to close the gap between body models used in computer vision and the anatomically accurate skeletons used in biomechanics. It builds BioAMASS, a dataset pairing over nine hours of SMPL body motion with fitted anatomical skeletons, by painting virtual markers on the skin and solving for the skeleton underneath. From this pairing it learns SKEL, a parametric model in which one set of shape and pose parameters $(\\boldsymbol{\\beta}, q)$ drives both the body surface and the bones, with anatomically constrained spine, shoulder-blade, and forearm motion. If the paper is right, the many existing systems that already estimate SMPL bodies can be upgraded to also output biomechanical joint locations and skeleton poses, enabling biomechanics from ordinary video and archived motion data.","feed_headline":"SKEL puts a real skeleton inside every SMPL body","feed_subtitle":"Same controls drive skin and bones, so vision datasets can gain anatomical joints.","key_machinery":"The load-bearing mechanism is a three-stage pipeline. First, a custom 24-bone biomechanical skeleton with 46 degrees of freedom—featuring a constant-curvature spine, an ellipsoid-sliding scapula, and a radius–ulna forearm—is fit inside posed SMPL meshes using virtual markers and a recent bilevel biomechanical optimization, producing the paired BioAMASS dataset. Second, a non-negative least-squares regressor maps SMPL mesh vertices to the 24 anatomical joint locations. Third, each bone's orientation is decomposed into a fixed base rotation learned from the dataset and a shape-dependent correction $R_{\\boldsymbol{\\beta}}^i(\\boldsymbol{\\beta})$ that aligns the bone with its parent–child joint segment; these pieces re-rig the SMPL mesh so the biomechanical pose parameters move skin and skeleton in synchrony. The decomposition is what lets SKEL keep SMPL's shape space while giving the bones anatomically coherent orientations inside every body shape.","core_discovery":"SKEL is presented as the first model in which the body surface and the anatomical skeleton are directly controlled by the same shape and pose parameters. Trained on the BioAMASS dataset, SKEL places a properly scaled skeleton inside any SMPL body; its regressed joint locations are closer to functional anatomical joints than SMPL's artist-defined joints, and its bones fit inside the body surface better than earlier skeleton-from-skin methods. The model uses fewer pose parameters than SMPL (46 versus 72) but models the spine as a constant-curvature chain, the scapula as sliding on an ellipsoid around the thorax, and forearm pronation and supination as a radius–ulna motion, so its articulation is biomechanically constrained rather than a collection of ball joints.","pith_inferences":["The decisive test the paper does not run is against independent anatomy: registering SKEL's femur-head and humerus-head joints to CT or MRI-derived joint centers on the same subjects would show whether the pseudo-ground-truth fits are unbiased.","If SKEL is used to train video-based biomechanics estimators, the natural downstream check is whether joint angles and moments computed from SKEL agree with marker-based motion capture to within clinical tolerances; that is the metric that would justify “biomechanics in the wild.”","Since the learning and rigging pipeline is independent of the particular skeleton model, the same procedure could be rerun with a clinically validated skeleton the moment one exists, upgrading SKEL without architectural changes."],"forward_implications":["Existing SMPL-based vision pipelines can be post-processed by fitting SKEL, turning datasets such as 3DPW and BEDLAM into sources of biomechanical skeleton parameters.","Regressing anatomical joints from skin (or fitting SKEL) yields joint locations closer to the functional joints than SMPL's own joints, with errors below one centimeter for most joints in the reported evaluation.","Because SKEL has 46 biomechanical degrees of freedom instead of SMPL's 72 ball-joint degrees, posed bodies exhibit more realistic shoulder-blade sliding, spine bending, and forearm twist.","Given a biomechanical skeleton, SKEL can generate a plausible skin surface while preserving the skeleton's bone lengths, allowing motion-capture data to be visualized with body shapes of different weights."],"supporting_citations":[{"why":"Defines the SMPL surface model whose mesh, shape space, and skinning weights SKEL inherits as the skin side of the paired model.","marker":"[Loper et al. 2015]"},{"why":"Provides the AMASS motion sequences of SMPL bodies that BioAMASS is built from.","marker":"[Mahmood et al. 2019]"},{"why":"Contributes the bilevel optimization used to fit the BSM skeleton to virtual markers, producing the pseudo-ground-truth joint locations.","marker":"[Werling et al. 2022]"},{"why":"OSSO's predicted bone geometry is used to personalize marker offsets on the skeleton and prevent bone over-stretching for large body shapes.","marker":"[Keller et al. 2022]"},{"why":"Supplies the scapulothoracic joint model with the scapula sliding on an ellipsoid, which BSM adapts for the shoulder.","marker":"[Seth et al. 2016]"},{"why":"Supplies the lower-body musculoskeletal model, including knee flexion, that BSM uses for the legs.","marker":"[Rajagopal et al. 2016]"},{"why":"Defines the musculoskeletal modeling framework in which the BSM skeleton and its bone meshes are authored.","marker":"[Delp et al. 2007]"},{"why":"DFAUST provides the held-out dataset of scan-fitted SMPL meshes used to evaluate joint regression and SKEL fit error.","marker":"[Bogo et al. 2017]"}],"fun_headline_variants":["SKEL re-rigs SMPL with real bone joints","Anatomical skeletons now drive SMPL bodies","Fewer parameters, true joints: SKEL model","Biomechanics in the wild with SKEL"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole result rests on the assumption that fitting the skeleton to markers painted on the skin of synthetic bodies yields essentially correct joint locations, even though those fits are only pseudo-ground-truth and are never checked against medical images or bone-pin measurements.","fun_headline_variants_meta":{"raw":{"variants":["SKEL re-rigs SMPL with real bone joints","Anatomical skeletons now drive SMPL bodies","Fewer parameters, true joints: SKEL model","Biomechanics in the wild with SKEL"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000243,"raw_usage":{"total_tokens":1574,"prompt_tokens":1033,"completion_tokens":541,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":649,"completion_tokens_details":{"reasoning_tokens":479}},"tokens_in":649,"tokens_out":541,"duration_ms":5044,"temperature":1.0,"reasoning_tokens":479,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:14:27.524723+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a subject with both a CT or MRI scan and synchronized marker motion capture, fit SKEL to the motion, and compare its femur-head and humerus-head joint centers with the imaging-derived centers; a systematic offset, especially at the shoulder, would falsify the claimed biomechanical accuracy.","supporting_citations":[],"review_version":2}