Pith. sign in

REVIEW 3 major objections 5 minor 24 references

BODIESReg registers 3D body scans to parametric body models by first aligning the template to the scan pose, achieving sub-10 mm mean surface-fit error on both test datasets.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 23:15 UTC pith:62GEUASL

load-bearing objection Useful, honest open-source registration pipeline; trust the code and the ablation, not the sub-10mm headline. the 3 major comments →

arxiv 2607.15463 v1 pith:62GEUASL submitted 2026-07-16 q-bio.QM q-bio.OT

BODIESReg: An Open-Source Pipeline for Registering 3D Body Scans Using Pose-Aligned Initialization

Vikash Chaurasia , Judit Cueto Fernandez , J. Micah Prendergast , Eline van der Kruk This is my paper
classification q-bio.QM q-bio.OT
keywords 3D body scan registrationparametric body modelpose-aligned initializationinverse kinematicsChamfer distancesurface-fit errorbiomechanical model personalizationbatch registration
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

BODIESReg claims that the main obstacle to registering raw 3D body scans to parametric body models is not the surface-distance optimization but the pose used to initialize it. The pipeline estimates the scan's pose from sparse keypoints, instantiates a pose-aligned template, and only then runs surface-distance and vertex-level fitting. On 746 synthetic close-interaction scans it succeeded automatically in 82.9% of cases; on 167 real optical scans it succeeded in all cases, with mean per-segment surface-fit error below 10 mm on both. The authors further claim this is the first fully automatic, open-source, dataset-training-free pipeline of this kind. If true, it makes large-scale subject-specific biomechanical model personalization practical.

Core claim

The paper's central claim is that the dominant failure mode in scan-to-body-model registration is poor pose initialization, not distance optimization. BODIESReg detects sparse keypoints from two orthographic projections of the aligned scan, solves inverse kinematics to drive the template into the scan's pose, and only then minimizes surface-distance (Chamfer) and vertex-level point-to-point error. On 746 synthetic close-interaction scans (CHI3D), 82.9% registered automatically; on 167 real optical scans (MorphoMotion), 100% registered, with mean per-segment surface-fit error below 10 mm in both datasets. The authors also claim this is the first published open-source, fully automatic, dataset

What carries the argument

The load-bearing mechanism is pose-aligned initialization: rather than starting from a T-pose template, the pipeline estimates the scan's 3D pose from sparse keypoints reconstructed from two orthographic projections using a pre-trained 2D landmark detector, then applies an inverse-kinematics decoder trained on human motion capture to produce a plausible posed template. This template seeds a regularized symmetric Chamfer-distance minimization, followed by injective (dense) or proximity-based (sparse) correspondence rematching and a final point-to-point vertex fit.

Load-bearing premise

The whole pipeline's anatomical correctness rests on the assumption that 2D keypoints detected in two orthographic projections of the scan—often textureless—reconstruct into 3D keypoints close enough for pose estimation to land the template where surface fitting converges to the anatomically correct solution; the paper itself reports this assumption failing on 128 of 746 synthetic scans.

What would settle it

Register a held-out set of textureless, extreme-pose scans with known ground-truth skeletons: the central claim predicts roughly an 83% automatic success rate with sub-10 mm mean error on the successes, and near-total failure if the pose-aligned initialization is replaced by a default T-pose template.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If correct, automatic batch registration of hundreds of body scans becomes feasible without per-dataset training or manual landmarking, enabling subject-specific biomechanical model personalization at scale.
  • Sub-10 mm mean surface-fit error on both synthetic and real scans supports use for automatic landmark extraction and segment parameter estimation, with the caveat that SMPL joint centers are surface-regressed rather than anatomical.
  • The ablation shows pose-aligned initialization is the decisive component: removing it drops success from 82.9% to 9.7% on CHI3D and from 100% to 71.9% on MorphoMotion.
  • Interactive fallbacks (pose editor, correspondence selector) extend the pipeline to scans where automatic keypoint detection fails, including MRI-derived supine skin surfaces, though with larger reported errors.
  • Compatibility with SMPL, SMPL+H, SMPL-X, and DMPL means the pipeline generalizes across parametric body models.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A testable extension: replacing the 2D landmark detector with one trained on depth maps or textureless renders could close the remaining CHI3D failure gap; the paper's own analysis implicates detector failure on textureless input.
  • The reported surface-fit error is an in-sample fitting residual measured on correspondences the pipeline itself establishes; anatomical correctness is adjudicated separately, so the sub-10 mm figure should not be read as independent validation of correspondence accuracy.
  • The left/right swap correction and the manual fallback tools indicate that pose initialization remains the fragile link; a pose-initialization module using full-surface cues (silhouette or geodesic features) rather than sparse keypoints would likely be more robust.
  • If the pose-first strategy generalizes, it suggests that other non-rigid registration problems with strong shape priors could benefit from the same two-stage split: estimate the latent configuration first, then refine with a correspondence-free distance.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper presents BODIESReg, an open-source pipeline for registering 3D body scans to SMPL-family parametric body models. The method first estimates a pose-aligned initial template from MediaPipe keypoints detected on two orthogonal projections, then refines the fit via regularized Chamfer distance minimization followed by correspondence rematching and point-to-point fitting. The authors report automatic registration success on 82.9% of CHI3D scans and on all 167 MorphoMotion scans, with mean surface-fit errors below 10 mm among successful registrations. They also describe an ablation showing that pose-aligned initialization markedly improves success rates, and they provide interactive tools for manual fallback. The paper emphasizes that it is the first published open-source framework for fully automatic, dataset-training-free scan-to-SMPL registration.

Significance. If the accuracy claims are reliable, BODIESReg would be a practically useful tool for biomechanical model personalization: it is open-source, has a reproducible CHI3D selection script, and the ablation demonstrates a clear and substantial benefit of pose-aligned initialization (CHI3D success rising from 9.7% to 82.9%). The optimization machinery is standard and the code availability is a genuine strength. However, the central accuracy evidence is weaker than the abstract suggests: the reported surface-fit errors are in-sample residuals computed on correspondences produced by the fitting itself, and the 'all MorphoMotion scans successful' claim rests on visual inspection rather than independent anatomical validation, despite whole-body MRI being available in the same dataset. The CHI3D success rate is also defined by a manually selected failure boundary. These issues do not invalidate the methodological contribution, but they require either additional validation or substantially tempered claims.

major comments (3)
  1. [Section 4.2.2, Eq. (14), Tables 1 and 2] The per-vertex surface-fit error is computed using p_scan_j = S[C(j)], where C(j) is the correspondence map established during the fitting procedure itself. Minimizing Eq. (14) directly reduces this quantity, so the reported mean errors are in-sample residuals of the optimization objective, not an independent measure of registration accuracy. The paper acknowledges this in Section 4.2.2, but the abstract's statement that 'mean surface-fit error are below 10 mm' for both datasets is still presented without this caveat. Because the paper's motivation is biomechanical model personalization, the distinction between surface fit and anatomical correctness is central. I recommend that the abstract and conclusions state explicitly that the <10 mm figures are fitting residuals, and that independent validation (e.g., against MRI-derived geometry or manually placed anatomical landmarks) is needed b
  2. [Section 4.2.1, Fig. 4] The CHI3D success rate of 82.9% depends on the visually selected failure boundary at rank 619 in the sorted emax curve. While the paper explicitly avoids an ad hoc numerical threshold, the boundary is still chosen by visual inspection, and no sensitivity or inter-rater information is provided. Since CHI3D provides ground-truth SMPL-X joints, a quantitative decision rule is feasible: for instance, report the success rate as a function of an emax threshold (e.g., 50, 100, 150 mm) or provide an ROC-style curve. This would make the central '82.9% success' claim reproducible and less dependent on one observer's judgment.
  3. [Sections 4.1.2 and 4.2.1] For MorphoMotion, no ground-truth joint positions are available and the paper states that 'we assess their anatomical correctness by visual inspection.' The claim that automatic registration succeeded for all 167 MorphoMotion scans is therefore based on subjective inspection and on in-sample surface residuals. The dataset contains whole-body MRI scans of the same participants, which could provide an independent, modality-different validation of skeletal geometry and joint placement, but the only MRI result is a single manual-fallback registration (Table 3) in which 10 of 24 segments exceed 10 mm mean error. I recommend either performing MRI-based validation on a subsample or explicitly reframing the MorphoMotion result as 'all scans passed visual inspection and achieved low fitting residuals' rather than as verified anatomical success.
minor comments (5)
  1. [Abstract] Grammar: 'mean surface-fit error were below 10 mm' should be 'mean surface-fit error was below 10 mm.'
  2. [Table 2 caption] The NA entries for sit-pose hips/upper legs/feet are explained in the text, but the caption could state explicitly that NA denotes regions where scan data were removed during chair/footrest cleaning, to avoid ambiguity for readers who scan tables only.
  3. [Appendix D.1] Typo: 'V olumes' should be 'Volumes' in the sentence 'These quantities are commonly used in inertial parameter estimation...'.
  4. [Section 3.1.3 and Appendix A] The derivation of SMPL keypoints from MediaPipe landmarks (e.g., spine interpolation, pelvis as hip midpoint) is plausible but would benefit from a figure or diagram showing the converted skeleton overlaid on a sample scan, beyond the skeleton-only illustration in Figure 9, to help readers judge the sensitivity of pose initialization to these heuristics.
  5. [Section 4.3.1 and Fig. 4] The inset in Figure 4 is described in the text as showing ranks 600-650, but the x-axis labels appear to be 600, 619, 650; please add axis titles and explain the shaded or marked region explicitly, since this figure is the basis for the central success-rate claim.

Circularity Check

2 steps flagged

Headline surface-fit accuracy is the in-sample residual of the optimized objective, and the MorphoMotion ablation uses the full pipeline's own output as pseudo-ground truth; the paper discloses both, but the central sub-10 mm claim is partly self-referential.

specific steps
  1. self definitional [Equation (14), Section 3.3.2; Section 4.2.2; Abstract]
    "Equation (14) minimizes Evtx = (1/3N) Σ_j ||v_j − p^scan_j||^2. Section 4.2.2: 'We define the per-vertex surface-fit error as ||v_j(β**,θ**)−p^scan_j|| and report the mean ± standard deviation... Because the scan-surface points are selected from correspondences established during fitting, this measure is an in-sample fitting residual and does not independently validate anatomical correspondence.' Abstract: 'mean surface-fit error were below 10 mm across both datasets.'"

    The reported 'surface-fit error' is exactly the per-vertex residual of the objective already minimized to obtain (β**,θ**) in Eq. (14), evaluated on the same correspondence map C(j) used in that optimization. The sub-10 mm headline is therefore a training loss, not an independent accuracy or anatomical metric. The paper's own sentence labels it 'an in-sample fitting residual'; presenting it as the main quantitative result makes the headline number self-referential by construction.

  2. other [Appendix C (Ablation Study)]
    "Ground-truth joint locations are unavailable for MorphoMotion. We therefore defined reference joints from the full automatic pipeline: each MorphoMotion scan was registered with the full pipeline, these registrations were visually verified as anatomically correct, and the corresponding SMPL parameters were passed through the SMPL forward model to obtain joint locations."

    For the MorphoMotion ablation, the pseudo-ground-truth emax reference is generated by the full pipeline being evaluated. Ablated registrations are therefore scored by distance to the with-initialization output, which favors the full pipeline by construction. Although final success classification is visual, this self-referential reference is used to sort and locate transition regions, so the reported ablation improvement on MorphoMotion is not an independent benchmark.

full rationale

The paper is largely self-contained and transparent; unlike a hidden fit-called-prediction, Section 4.2.2 explicitly admits the surface-fit metric is an in-sample residual, and Section 5.1 states SMPL joint centers are not biomechanically correct. However, the abstract and headline 'mean surface-fit error below 10 mm' still present this residual as the main accuracy result, and the MorphoMotion ablation uses full-pipeline output as pseudo-ground-truth, so the central quantitative claims partially reduce to the fitting objective and to the method's own output. The MorphoMotion 'all scans successful' determination by visual inspection and the absence of MRI validation of the 167 automatic registrations are external-validity limitations rather than circular reductions, but they reinforce that no independent anatomical ground truth supports the strong claims. For CHI3D, failure classification uses ground-truth SMPL-X joints, providing some independent content; this prevents the score from reaching 8-10. Overall score 6.

Axiom & Free-Parameter Ledger

8 free parameters · 6 axioms · 0 invented entities

The pipeline's numerical outputs depend on several hand-set weights and thresholds (λ's, α, dense/sparse threshold, search-box geometry, visual failure boundary). No new physical entities are introduced; the assumptions are domain assumptions about the fidelity of SMPL, VPoser, MediaPipe, and the correspondence model.

free parameters (8)
  • VPoser regularization weights λz, λβ = 0.001
    Set empirically below VPoser defaults (0.01, 0.5) to allow pose freedom (Section 3.1.4).
  • IK data weight α = 20
    Increased from default 10 to make keypoint matching dominant (Section 3.1.4).
  • Chamfer regularization weights λβ, λθ = 0.001
    Weak regularization anchored to IK estimates, used for all experiments (Section 3.2).
  • Adam learning rate and iteration count = 0.005, 6000 iters
    Selected for convergence; no early stopping (Section 4).
  • Dense/sparse scan threshold = 30,000 vertices
    Empirically chosen as the point where injective matching stayed stable (Section 3.3.1).
  • Correspondence search box initial half-width and expansion = r=h/50, factor 1.5, max 5 searches
    Empirically selected during method development (Section 3.3.1).
  • Number of keypoint joints = 22 (exclude 2 hand joints)
    Hand joints not reliably detected from projections (Section 3.1.3).
  • CHI3D failure boundary rank = 619
    Visually selected transition in sorted emax; determines the 82.9% success rate (Section 4.2.1).
axioms (6)
  • domain assumption SMPL-family body models faithfully represent human shape and pose for the scanned populations
    The pipeline treats SMPL/SMPL-X/SMPL+H/DMPL as ground-truth geometry generators; cited from [9,10,20].
  • domain assumption VPoser latent space covers the poses in the test scans and L-BFGS reaches a valid pose
    Adapted from Pavlakos et al. [10]; relied on in Eq. (8).
  • domain assumption MediaPipe keypoint detection transfers to rendered orthographic projections of 3D scans/meshes
    Core to pose initialization; acknowledged as unreliable on textureless scans (Section 5.1.1).
  • domain assumption PCA-based coordinate alignment and the left/right heuristic (Appendix B) canonicalize the scan correctly
    If alignment or handedness is wrong, keypoint reconstruction fails (Section 3.1.1).
  • domain assumption Nearest-neighbor (Chamfer) correspondences are anatomically correct once the pose-aligned initialization is close
    The method's operating assumption; mitigations via interactive correspondence selector (Section 3.3.1).
  • domain assumption CHI3D ground-truth joints from SMPL-X forward pass are error-free
    Used to compute emax and failure classification (Section 4.2.1).

pith-pipeline@v1.3.0-alltime-deepseek · 21771 in / 13016 out tokens · 119962 ms · 2026-08-01T23:15:29.831380+00:00 · methodology

0 comments
read the original abstract

Biomechanical models are used to quantify and optimize human movement in clinical rehabilitation, sports science, and occupational health. Personalizing these models requires accurate identification of anatomical landmarks and body segment parameters, which can be derived from 3D body surface scans. Registering these surfaces to a canonical template is essential for automated landmark detection and model scaling. However, fitting parametric body models to real-world 3D point clouds remains challenging: non-linear optimization can converge to suboptimal solutions when target poses deviate substantially from the default pose of a template. We present BODIESReg, an open-source registration pipeline that addresses this initialization problem by constructing a pose-aligned mesh before distance minimization begins. In automatic mode, BODIESReg can be used for end-to-end registration without user intervention and processes multiple scans in batch. We evaluated BODIESReg on two complementary datasets: CHI3D, a synthetic dataset containing complex human poses, and MorphoMotion, a set of real-world 3D optical scans. Automatic registration succeeded for 82.9% of CHI3D scans and for all MorphoMotion scans. Among successful registrations, mean surface-fit error were below 10 mm across both datasets. For cases where automatic initialization fails, we provide interactive tools for manual pose correction and correspondence selection. BODIESReg supports large-scale registration of 3D body scans for biomechanical research.

Figures

Figures reproduced from arXiv: 2607.15463 by Eline van der Kruk, J. Micah Prendergast, Judit Cueto Fernandez, Vikash Chaurasia.

Figure 1
Figure 1. Figure 1: Key steps of the BODIESReg registration pipeline. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Interactive pose editor. Left: a 3D point cloud overlaid with the default 22-joint SMPL skeleton (two hand [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Interactive correspondence selector used in manual mode. The user selects paired regions on the input 3D scan [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Maximum joint error emax for all 746 CHI3D registrations, sorted in ascending order. The inset shows the transition region (ranks 600–650) where emax rises sharply; registrations above rank 618 are classified as failures and excluded from accuracy evaluation. keypoint detection failure registration failure [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Illustration of the cascading failure in the CHI3D dataset. Left: front and side views of a standing subject; [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Representative examples of CHI3D registrations using the automatic pipeline. [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Representative scans from the MorphoMotion dataset and their registered point clouds (shown in red). From [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Representative MRI skin surface registration. Front and side views show the input skin surface (blue) and [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: MediaPipe and SMPL keypoint skeletons used in pose initialization. MediaPipe landmarks are shown on the [PITH_FULL_IMAGE:figures/full_fig_p018_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Maximum joint error emax, defined in (15), for all 746 CHI3D registrations performed without pose initialization, sorted in ascending order. The inset magnifies the transition region around the failure boundary. Because the error grows gradually rather than through a sharp jump, the boundary was located by manual inspection: registrations up to index 72 were verified to be anatomically correct, while thos… view at source ↗
Figure 11
Figure 11. Figure 11: Maximum joint error emax, defined in (15), relative to full-pipeline reference joints, for MorphoMotion registrations performed without pose initialization, sorted in increasing order for each pose and sex. Left: A-pose registrations exhibit an abrupt transition between correct and failed registrations. Right: sit-pose registrations exhibit a more gradual transition. Intermediate errors of 121–390 mm corr… view at source ↗
Figure 12
Figure 12. Figure 12: SMPL template segmented into 24 anatomical regions, each shown in a distinct color, with the underlying [PITH_FULL_IMAGE:figures/full_fig_p023_12.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

24 extracted references · 7 canonical work pages · 2 internal anchors

  1. [1]

    Sorkine, M

    O. Sorkine, M. Alexa, As-rigid-as-possible surface modeling, in: SGP ’07: Proceedings of the Fifth Eurographics Symposium on Geometry Processing, Eurographics Association, 2007, pp. 109–116. doi:10.2312/SGP/SGP07/109- 116

    work page doi:10.2312/sgp/sgp07/109- 2007

  2. [2]

    Yoshiyasu, W.-C

    Y . Yoshiyasu, W.-C. Ma, E. Yoshida, F. Kanehiro, As-conformal-as-possible surface registration, Computer Graphics Forum 33 (5) (2014) 257–267. doi:10.1111/cgf.12451

    doi:10.1111/cgf.12451 2014

  3. [3]

    Makadia, A

    A. Makadia, A. Patterson, K. Daniilidis, Fully automatic registration of 3D point clouds, in: CVPR ’06: Proceed- ings of the 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, V ol. 1, IEEE Computer Society, 2006, pp. 1297–1304. doi:10.1109/CVPR.2006.122

    work page doi:10.1109/cvpr.2006.122 2006

  4. [4]

    Amberg, S

    B. Amberg, S. Romdhani, T. Vetter, Optimal step nonrigid ICP algorithms for surface registration, in: 2007 IEEE Conference on Computer Vision and Pattern Recognition, IEEE, 2007, pp. 1–8. doi:10.1109/CVPR.2007.383165

    arXiv 2007

  5. [5]

    Ovsjanikov, M

    M. Ovsjanikov, M. Ben-Chen, J. Solomon, A. Butscher, L. Guibas, Functional maps: A flexible representation of maps between shapes, ACM Transactions on Graphics (TOG) 31 (4) (2012) 30. doi:10.1145/2185520.2185526

    arXiv 2012

  6. [6]

    Weakly Supervised Deep Functional Map for Shape Matching

    A. Sharma, M. Ovsjanikov, Weakly supervised deep functional maps for shape matching, Advances in Neural Information Processing Systems 33 (2020) 19264–19275. doi:10.48550/arXiv.2009.13339

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2009.13339 2020

  7. [7]

    Attaiki, G

    S. Attaiki, G. Pai, M. Ovsjanikov, DPFM: Deep partial functional maps, in: 2021 International Conference on 3D Vision (3DV), IEEE, 2021, pp. 175–185. doi:10.1109/3DV53792.2021.00040

    arXiv 2021

  8. [8]

    Anguelov, P

    D. Anguelov, P. Srinivasan, D. Koller, S. Thrun, J. Rodgers, J. Davis, SCAPE: Shape completion and animation of people, ACM Transactions on Graphics 24 (3) (2005) 408–416. doi:10.1145/1073204.1073207

    arXiv 2005

  9. [9]

    Loper, N

    M. Loper, N. Mahmood, J. Romero, G. Pons-Moll, M.J. Black, SMPL: A skinned multi-person linear model, ACM Transactions on Graphics 34 (6) (2015) 248:1–248:16. doi:10.1145/2816795.2818013

    arXiv 2015

  10. [10]

    Pavlakos, V

    G. Pavlakos, V . Choutas, N. Ghorbani, T. Bolkart, A.A. Osman, D. Tzionas, M.J. Black, Expressive body capture: 3D hands, face, and body from a single image, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019, pp. 10975–10985. doi:10.1109/CVPR.2019.01123

    arXiv 2019

  11. [11]

    Osman, T

    A.A. Osman, T. Bolkart, M.J. Black, STAR: Sparse trained articulated human body regressor, in: Computer Vision: ECCV 2020, 16th European Conference, Proceedings, Part VI, Springer-Verlag, 2020, pp. 598–613. doi:10.1007/978-3-030-58539-6_36

    work page doi:10.1007/978-3-030-58539-6_36 2020

  12. [12]

    Schmidt, I

    A.M. Schmidt, I. Peraza, Y . Kyosev, Evaluation of automated skeleton fitting to 4D human body scan data using open-source SMPL and OSSO models, International Journal of 3D Body Technologies 1 (2024) 40. doi:10.15221/24.40

    work page doi:10.15221/24.40 2024

  13. [13]

    F. Bogo, A. Kanazawa, C. Lassner, P. Gehler, J. Romero, M.J. Black, Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image, in: Computer Vision: ECCV 2016, Lecture Notes in Computer Science, V ol. 9909, Springer, Cham, 2016, pp. 561–578. doi:10.1007/978-3-319-46454-1_34

    doi:10.1007/978-3-319-46454-1_34 2016

  14. [14]

    Kolotouros, G

    N. Kolotouros, G. Pavlakos, M.J. Black, K. Daniilidis, Learning to reconstruct 3D human pose and shape via model-fitting in the loop, in: 2019 IEEE/CVF International Conference on Computer Vision (ICCV), IEEE, 2019, pp. 2252–2261. doi:10.1109/ICCV .2019.00234

    arXiv 2019

  15. [15]

    J. Li, C. Xu, Z. Chen, S. Bian, L. Yang, C. Lu, HybrIK: A hybrid analytical-neural inverse kinematics solution for 3D human pose and shape estimation, in: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), IEEE, 2021, pp. 3383–3392. doi:10.1109/CVPR46437.2021.00339

    arXiv 2021

  16. [16]

    LoopReg: Self-supervised Learning of Implicit Surface Correspondences, Pose and Shape for 3D Human Mesh Registration

    B.L. Bhatnagar, C. Sminchisescu, C. Theobalt, G. Pons-Moll, LoopReg: Self-supervised learning of implicit surface correspondences, pose and shape for 3D human mesh registration, Conference on Neural Information Processing Systems (NeurIPS) 33 (2020) 12909–12922. doi:10.48550/arXiv.2010.12447

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2010.12447 2020

  17. [17]

    H. Feng, P. Kulits, S. Liu, M.J. Black, V . Fernandez Abrevaya, Generalizing neural human fitting to unseen poses with articulated SE(3) equivariance, in: 2023 IEEE/CVF International Conference on Computer Vision (ICCV), IEEE, 2023, pp. 7977–7988. doi:10.1109/ICCV51070.2023.00733 24 BODIESRegA PREPRINT

    arXiv 2023

  18. [18]

    Lugaresi, J

    C. Lugaresi, J. Tang, H. Nash, et al., MediaPipe: A framework for building perception pipelines, arXiv preprint arXiv:1906.08172 (2019). doi:10.48550/arXiv.1906.08172

    Pith/arXiv arXiv doi:10.48550/arxiv.1906.08172 1906

  19. [19]

    Barrow, J.M

    H.G. Barrow, J.M. Tenenbaum, R.C. Bolles, H.C. Wolf, Parametric correspondence and chamfer matching: Two new techniques for image matching, in: IJCAI’77: Proceedings of the 5th International Joint Conference on Artifi- cial Intelligence, V ol. 2, Morgan Kaufmann Publishers Inc., 1977, pp. 659–663. doi:10.5555/1622943.1623000

    arXiv 1977

  20. [20]

    Romero, D

    J. Romero, D. Tzionas, M.J. Black, Embodied hands: Modeling and capturing hands and bodies together, arXiv preprint arXiv:2201.02610 (2022). doi:10.48550/arXiv.2201.02610

    Pith/arXiv arXiv doi:10.48550/arxiv.2201.02610 2022

  21. [21]

    tudelft.nl/dhpc/ark:/44463/DelftBluePhase2

    Delft High Performance Computing Centre (DHPC), DelftBlue Supercomputer (Phase 2), 2024, https://www. tudelft.nl/dhpc/ark:/44463/DelftBluePhase2

    2024

  22. [22]

    Fieraru, M

    M. Fieraru, M. Zanfir, E. Oneata, A.-I. Popa, V . Olaru, C. Sminchisescu, Three-dimensional reconstruction of human interactions, in: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), IEEE, 2020, pp. 7214–7223. doi:10.1109/CVPR42600.2020.00724

    arXiv 2020

  23. [23]

    Fieraru, M

    M. Fieraru, M. Zanfir, E. Oneata, A.-I. Popa, V . Olaru, C. Sminchisescu, Reconstructing three-dimensional models of interacting humans, IEEE Transactions on Pattern Analysis and Machine Intelligence 47 (12) (2025) 10870–10881. doi:10.1109/TPAMI.2025.3601974

    arXiv 2025

  24. [24]

    Cueto Fernandez, C

    J. Cueto Fernandez, C. van de Steeg-Henzen, A.C. Schouten, A. Seth, E. van der Kruk, MSK-Morph: An automated framework to systematically morph landmark-defined musculoskeletal models into subject-specific bone geometries, bioRxiv (2024). doi:10.64898/2026.04.28.721455 25

    work page doi:10.64898/2026.04.28.721455 2024