acceptodds
Under review as a conference paper at ICLR 2027

SPUD: Skeletal Physics-based human pose estimation Using Diffusion

Abstract

We propose SPUD as a diffusion-based framework for physics-based human pose estimation (PHPE), regularized by a novel kinetics prior and a novel SKEL-based pose prior. We also present a fencing motion-capture video dataset, synchronized with force plates, to allow for quantitative evaluation of kinetics estimation for athletic human movements. PHPE methods aim to improve realism by enforcing the laws of dynamics. However, inverse dynamics optimization without force annotations is challenging, as it is an underconstrained problem. To address this, we turn to diffusion models and their expressiveness in learning complex distributions, as recent studies have demonstrated impressive results for solving inverse problems via variational diffusion sampling. We build on these frameworks for inverse kinematics, and extend them to inverse dynamics by training a novel kinetics prior on simulation data. Furthermore, for improved compatibility with downstream biomechanics analysis, SPUD is the first PHPE method to infer human dynamics for the SKEL human model. Lastly, we conduct extensive experiments on various inverse kinematics and inverse dynamics tasks to demonstrate SPUD's versatility, and on datasets with disparate human movements to demonstrate improved adaptability. This includes a force-annotated fencing dataset we collected, which can facilitate more robust evaluations for future PHPE research.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
human pose estimation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_rHvJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents SPUD, a framework for physics-based human pose estimation that combines diffusion priors with test-time optimization. Using the SKEL body model, SPUD incorporates a pose prior, DBone, and a kinetics prior, DForce, to regularize human motion and ground reaction forces. It optimizes pose, anthropometric parameters, and contact forces using these priors and a dynamics residual loss. The authors also collect a fencing dataset with synchronized video, motion capture, and force plate measurements for quantitative force evaluation. Experiments cover general human movements, fencing, golf, and motion denoising, showing improvements in overall vertical force error and selected motion plausibility metrics, although the benefits vary across tasks and metrics.

Critical strengths
  1. A well-motivated approach to a practically important ambiguity in dynamics estimation. Estimating contact forces from monocular human motion is underconstrained, so regularizing the solution with a learned joint prior over motion and forces is a reasonable approach. Combining this idea with SKEL and a dynamics model that supports optimization has the potential to connect visual human reconstruction with biomechanical analysis. This is the paper's most valuable methodological direction.

  2. Real force plate measurements and measurable empirical improvements. The fencing dataset provides synchronized video, motion capture, and force plate measurements, enabling direct evaluation of predicted forces. In Table 2, SPUD reduces overall vertical force error from MusclePose's 0.276 to 0.249, an improvement of approximately 9.8%, while reducing the proportion of floating frames from 37.6% to 6.8%. These results demonstrate benefits on this test set and illustrate the potential value of the dataset.

Critical weaknesses
  1. The core additional contribution is not sufficiently validated. SPUD builds on DPoser-X's optimization framework, with its methodological novelty primarily residing in DForce and its integration with SKEL dynamics. However, the main experiments compare only against two regression-based methods, and the w/o prior ablation does not isolate DForce, leaving the additional value of the force diffusion prior unclear.

  2. The trade-off between pose accuracy and physical plausibility is insufficiently supported. In Table 1, SPUD achieves the lowest foot-sliding error, but its H36M-Walk PJE is 49.5 mm, compared with 41.0 mm for PhysPT and 42.9 mm for MusclePose, and other contact metrics do not consistently improve. The force quantiles lack ground-truth comparisons and cannot establish greater force accuracy. The only evaluation with measured forces, fencing, includes two subjects and acknowledged noise and undermeasurement due to gaps between force plates. The current results therefore do not sufficiently establish the framework's kinetic benefits and associated costs in pose accuracy.

Questions for authors
  1. What specifically distinguishes DForce from existing motion-force diffusion models? Could you report pose and force errors with only DForce removed, keeping all other settings fixed, to establish its independent contribution?

  2. How do you explain the pose accuracy degradation on H36M-Walk? Does the improvement in fencing force estimation over MusclePose persist when evaluation is restricted to valid force plate measurement intervals?

5 comments

Dear Reviewer rHvJ. Thank you for taking the time to review our paper. We have updated the paper accordingly based on the Weaknesses (Ws) and Questions (Qs) you've posed, addressed individually below. Your suggested experiments were added to Table 2 and a new Table 6 (page 16) of the revised version of the paper, with the corresponding results sections also updated.


(W1) On limited ablations to demonstrate the effects of the priors.

Thank you for this suggestion. In Table 2 and the new Table 6, we have added ablations to demonstrate how the individual priors contributed separately to the kinematic and kinetic improvements. The discussion in the results sections were also updated correspondingly. We have also moved the motion denoising experiments from the appendix to the last page of the main paper, to show the contributions of DBone.


(W2) On the current results not sufficiently establishing the framework's kinetic (W2b) benefits and associated costs in pose accuracy (W2a).

We address our kinematics (W2a) and kinetics (W2b) contributions separately below. As discussed in paragraph 3 of Sec 4.1, our comparisons are mainly with 2 recent physics-based human pose estimators, PhysPT and MusclePose, because the former produces low pose errors (corresponds to W2a) while the latter produces more plausible kinetic predictions (corresponds to W2b).

(W2a) The trade-off between pose accuracy and physical plausibility is insufficiently supported. In Table 1, SPUD achieves the lowest foot-sliding error, but its H36M-Walk PJE is 49.5 mm, compared with 41.0 mm for PhysPT and 42.9 mm for MusclePose, and other contact metrics do not consistently improve.

  • in terms of positional accuracy, among the five evaluation datasets, our method only performed worse than MusclePose on one (H36M-walk, discussed in Q2), and was significantly better than MusclePose on the other four, slightly behind PhysPT. This lag behind PhysPT may also be due to an inherent fitting bias as SKEL has fewer degrees of freedom (less flexible) than SMPL, as discussed in Sec 4.1 paragraph 2 and in [1].

  • overall, we also showed improvements in common kinematic plausibility metrics such as floating, ground penetration, and sliding.

  • to your last point, we have also added a more detailed analysis in Tab 6, to show improvements in contact estimation.

(W2b) The force quantiles lack ground-truth comparisons and cannot establish greater force accuracy. The only evaluation with measured forces, fencing, includes two subjects and acknowledged noise and undermeasurement due to gaps between force plates.

  • You are correct. One of the problems with existing literature in physics-based HPE is the lack of quantitative evaluation on the inferred forces, as the main goal of incorporating dynamics in previous works is often to reduce kinematic artifacts, and also due to a lack of force-annotated video motion datasets to evaluate on. Hence our motivation to collect this fencing dataset to quantitatively evaluate the kinetics of athletic movements, for which our method showed improvements over MusclePose in both joint position errors and force errors.

  • among the other four evaluation datasets without force-annotations, our method was similar to MusclePose in terms of kinetic plausibility.

  • to your second point, for more reliable force evaluation, in Table 6, we further analyzed the results for the frames where the force errors were isolated from potential contact errors, and showed that our method also outperformed others in this case. Following this analysis, we also believe certain low force errors produced by MusclePose were actually due to their worse contact estimation.

  • for robustness, results in Tab 6, were grouped by different subjects and movements. MusclePose's results (good for the novice fencer S2 and poor for the expert fencer S1) seemed to suggest a subject effect, whereas our results were more consistent across the two subjects and different scenarios, indicating better generalizability towards the better and faster fencer (S1). And in the area of most practical significance — the back foot of a lunge, MusclePose’s force errors for the expert’s contact frames were almost twice of the novice, while our method improved MusclePose’s by about 44%.

  • furthermore, your suggested ablations from W1 also show how DForce contributes to these kinetic improvements above. As the force errors were significantly higher in the rows with DForce removed.

  • lastly, in Sec A.1, we added a discussion on why we mainly reported force errors in the vertical direction and why we believe it is reliable. For example, in comparison, the shear/horizontal readings may not be as reliable, as we taped anti-slip mats onto the force plates for participant safety. As discussed in [2], the effects of force plate coverings on the horizontal components can be much larger than the vertical component, and with the vertical component mainly affected due to material thickness/cushioning/impact absorption. In our case, the anti-slips mats were very thin (small effect on vertical components), and had to be frequently replaced due to tearing and bunching (adds shear compliance, potentially affecting horizontal readings).

[1] Xia et al, “Reconstructing humans with a biomechanically accurate skeleton,” CVPR, 2025.
[2] Smith and Ditroilo, “Force plate coverings significantly affect measurement of ground reaction forces,” PLoS ONE, 2023.


(Q1a) What specifically distinguishes DForce from existing motion-force diffusion models?

Most diffusion-based human pose estimators [3,4] in general follow a multi-hypothesis approach that generates multiple 3D pose candidates to handle ambiguity. i.e. they use a conditional diffusion model to approximate the posterior distribution, and from which sample from. Our method extends the variational inference framework in [5,6] to inverse dynamics, using diffusion models as unconditional priors for regularization, to maximize evidence/marginal likelihood (lower bound). This provides more flexibility as it is not tailored for specific generation tasks, domains, or conditional inputs.

In terms of physics-based pose estimators, ours is also the first to directly incorporate SKEL-dynamics. This provides better compatibility with downstream biomechanics applications.

[3] Holmquist and Wandt, "Diffpose: Multihypothesis human pose estimation using diffusion models," ICCV, 2023.
[4] Cho and Kim, "Generative approach for probabilistic human mesh recovery using diffusion models." ICCV, 2023.
[5] Mardani et al, "A variational perspective on solving inverse problems with diffusion models," ICLR, 2024.
[6] Lu et al, "DPoser-X: Diffusion model as robust 3d whole-body human pose prior," CVPR, 2025.

(Q1b) Could you report pose and force errors with only DForce removed, keeping all other settings fixed, to establish its independent contribution?

Yes. As discussed in W1, we have added an experiments to isolate DForce's contributions to the kinetics improvements in Tables 2 and 6.


(Q2a) How do you explain the pose accuracy degradation on H36M-Walk?

We believe this may be partially due to the lower frame rate, as we followed common protocol to evaluate H36M at 10 fps, whereas the other evaluation sets were at 30 fps. And positional accuracy may be improved here by further tuning the temporal smoothing weight in the auxillary loss. Another reason for this degradation may be caused by compensating for the error build up from the global trajectory estimation, as the total distance traveled per video is the largest in this evaluation set.


(Q2b) Does the improvement in fencing force estimation over MusclePose persist when evaluation is restricted to valid force plate measurement intervals?

Yes. As discussed in W2b, we have added Table 6 to show that our improvements persist and increase for more reliable error metrics, such as the vertical force error for contact frames.



Note: the VFElungeba\text{VFE}^{ba}_{lunge} values (force errors during a lunge) in Table 2 are slightly changed due to correcting a previous error when manually segmenting the lunge frames. This is also reflected in the bottom row of Figure 4, where the previous version had an erroneous curve with a "second" peak around 60-80% of the lunge, and is no longer apparent. However, this is insignificant for the overall evaluation and comparisons remain the same.

Reviewer_MJNUOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SPUD, a test-time optimization framework for physics-based human pose estimation (PHPE) that extends diffusion-prior-based inverse kinematics to inverse dynamics. The method is built on the biomechanically motivated SKEL human model and introduces two unconditional diffusion priors: DBone, a pose prior defined over SKEL joint rotations, and DForce, a kinetics prior designed to regularize ground reaction force estimation.

SPUD jointly optimizes pose, body shape, anthropometric parameters, and force-related variables under kinematic reconstruction, dynamics consistency, auxiliary, and diffusion-prior losses. The method is evaluated on standard human pose benchmarks including 3DPW, RICH, and Human3.6M, as well as on GolfDB and a newly collected fencing dataset synchronized with marker-based motion capture and force plates. The paper also evaluates DBone on motion denoising and occlusion tasks.

The main contributions are the extension of diffusion-prior-based test-time optimization from inverse kinematics to inverse dynamics, the use of SKEL for biomechanically compatible PHPE, the proposed DForce and DBone priors, and a new force-annotated fencing dataset for evaluating athletic human motion.

Critical strengths
  1. Extending the optimization framework based on diffusion prior to physics-based human motion estimation is a meaningful technical exploration. This paper addresses a core challenge in physics-based human pose estimation: the inverse dynamics problem is significantly under-constrained in the absence of direct mechanical measurements. The author introduces learned dynamic priors to regularize the force estimation process, which is a highly reasonable and targeted approach. In addition, using SKEL as the human body model aligns well with the biomechanical application goals emphasized in the paper. Compared to the conventional kinematic skeleton based on SMPL, SKEL is closer to the anatomical structure in terms of joint centers and skeleton definitions, making it more suitable for subsequent biomechanical analysis.

  2. The newly constructed fencing dataset with mechanical annotations has high evaluation value. Quantitative evaluation of human biomechanics, especially force estimation results, has always been limited by the lack of synchronized mechanical annotation data. The fencing dataset collected in this paper includes RGB videos, marker-based motion capture data, and force plate measurement results, covering relatively fast and dynamic sports. Compared to traditional walking or daily action benchmarks, this dataset provides a more challenging and valuable testing environment for evaluating models' kinematic and dynamic prediction capabilities in high-dynamic motion scenarios.

Critical weaknesses
  1. The core argument for improving dynamic estimation is supported by only limited and inconsistent quantitative evidence. Improving force estimation is one of the main motivations for SPUD, but the most compelling quantitative validation was conducted on a relatively small fencing dataset that included only two subjects. Furthermore, although the proposed model is capable of simulating a richer set of force and torque-related quantities, the paper primarily reports on vertical force errors.

  2. In the GolfDB experiment, the claimed "adaptability/generalization ability" is difficult to distinguish from the specific task constraints introduced during the testing phase. In the GolfDB experiment, SPUD explicitly introduced golf-specific constraints in the auxiliary objective function, including reducing foot sliding, keeping wrist distance close, and extending the front elbow. However, in the evaluation phase, the researchers used the degree of foot sliding, wrist distance, and elbow extension as the main indicators of success. Therefore, some of the performance improvements reported in the experiment actually reflect the optimization process during the testing phase being able to meet the artificially set constraints of the golf motion, rather than directly proving that the learned posture or dynamic priors have better generalization capabilities when facing unseen motions. Although the paper includes an ablation experiment that removes LgolfL_{\mathrm{golf}}, the results also indicate that the significant improvements in these metrics are largely driven by the specific task loss function introduced by humans. Therefore, I believe the current claims about "adaptability" are somewhat exaggerated; unless the authors can more clearly distinguish "optimization flexibility" from "prior generalization ability."

Questions for authors

Question 1: Can the authors provide an ablation study that examines DForce and DBone separately? The current "w/o prior" ablation experiment seems to have removed both priors simultaneously. Given that DForce is the core methodological contribution of this paper, at least the following comparisons should be made: (i) the complete SPUD model; (ii) removing DForce; (iii) removing DBone; (iv) simultaneously removing both priors. This will help clarify whether the reported improvements in dynamic performance are due to DForce or the overall optimization framework.

Question 2: How can the experimental results on GolfDB be interpreted as evidence of "prior generalization capability" rather than "optimization for a specific task during the test phase"? Several evaluation metrics in GolfDB directly correspond to the manually introduced LgolfL_{\mathrm{golf}} constraint. Could the authors clearly indicate which results demonstrate that DBone/DForce itself possesses generalization ability? Especially, it would be very valuable if the authors could provide a comparison of force-related results without using the LgolfL_{\mathrm{golf}} constraint.

Question 3: Can the authors provide a more detailed analysis of the force results in fencing at different levels of subjects or types of movements, and discuss their statistical reliability? The current dataset only includes two subjects, and the model's performance difference compared to MusclePose varies between professional fencers and beginners. If statistical data, uncertainty estimates, or further analysis can be provided for individual subjects or action types, it would help determine whether the observed performance improvement is robust or merely dominated by effects from specific subjects.

2 comments

Dear Reviewer MJNU. Thank you so much for your insightful feedback. We have updated the paper accordingly based on the Weaknesses (Ws) and Questions (Qs) you've posed, with a summary below. Results for the suggested ablation experiments are added to Table 2, and a more detailed analysis in Table 6 on Page 16.


(W1) On limited force evaluation.
You are correct. One of the problems with existing literature in physics-based HPE is the lack of quantitative evaluation on the estimated forces (eg [1,2]), as the main goal of incorporating dynamics in previous works was often to reduce kinematic artifacts. Hence our motivation to collect this force-annotated fencing dataset to evaluate the kinetics of athletic movements.

On limited kinetics error metrics reporting.
We have updated Table 2 to include magnitude errors for the ground reaction force (GRF) vector, and added a Table 6 (page 16) that includes errors for the non-vertical GRF components, to show that our method improved in these areas as well.

We focused on reporting errors for the GRFs because granular joint torques and forces are downstream outputs that can be computed analytically via inverse dynamics from the kinematics and GRFs. Furthermore, we previously focused on the vertical component because (i) it is often the largest component and often the component of interest in sports, and (ii) the shear/horizontal readings may not be as reliable, as we taped anti-slip mats onto the force plates for participant safety. As discussed in [3], the effects of force plate coverings on the horizontal components can be much larger than the vertical component, and with the vertical component mainly affected due to material thickness/cushioning/impact absorption. In our case, the anti-slips mats were very thin (small effect on vertical components), and had to be frequently replaced due to tearing and bunching (adds shear compliance, potentially affecting horizontal readings). We have updated the paper to include these explanations in Sec A.1 (page 16).

[1] Li et al, "PhysPT: Physics-aware pretrained transformer for estimating human dynamics from monocular videos," CVPR, 2024.
[2] Zhu et al., "3D Human pose estimation with muscles," NeurIPS, 2025.
[3] Smith and Ditroilo, “Force plate coverings significantly affect measurement of ground reaction forces,” PLoS ONE, 2023.


(W2, Q1, Q2) On clarification between "optimization flexibility" from "prior generalization ability", and ablations to demonstrate the latter.
Again, you are correct. The GolfDB experiment demonstrates more of "optimization flexibility" than "prior generalization ability". We have clarified the distinction between these two claims in the paper: in the 2nd last paragraph of Sec 1 (page 2), end of "fencing" paragraph and start of "golf" paragraph (page 8), and in the conclusion (Sec 5, page 10), we specified adaptability in terms of showing prior generalizability for fencing and optimization flexibility for golf.

In Table 2 and the new Table 6, we have added the ablations to demonstrate how the individual priors contributed separately to the kinematic and kinetic improvements. The discussion in the results sections were also updated correspondingly. We have also moved the motion denoising experiments from the appendix to the last page of the main paper, to show the contributions of DBone.


(Q3) On more detailed analysis of force results in fencing.

We added a Table 6 (page 16) that includes a more detailed evaluation of the fencing results grouped by subject, movements, and contact, to more robustly demonstrate SPUD's improvement in performance as well as the contributions of DForce.

Summary of these results include:

  • in addition to the vertical direction, we further showed that our method produced lower force errors in the forward direction. Error comparisons for lateral direction were insignificant in magnitude.
  • for more reliable force evaluation, we further showed lower errors after isolating the contact frames. We also showed better contact estimation, and deduced that certain low force errors from MusclePose were due to poor contact estimation.
  • MusclePose's results (good for the novice fencer S2 and poor for the expert fencer S1) did seem to suggest a subject effect, whereas our results were more consistent across the two subjects and different scenarios, indicating better generalizability towards the better and faster fencer (S1). And in the area of most practical significance — the back foot of a lunge, MusclePose’s force errors for the expert’s contact frames were almost twice of the novice, while our method improved MusclePose’s by about 44%.
  • ablation results are also included in this Table 6 to show the contributions of DForce in these areas.


Note: the VFElungeba\text{VFE}^{ba}_{lunge} values (force errors during a lunge) in Table 2 are slightly changed due to correcting a previous error when manually segmenting the lunge frames. This is also reflected in the bottom row of Figure 4, where the previous version had an erroneous curve with a "second" peak around 60-80% of the lunge, and is no longer apparent. However, this is insignificant for the overall evaluation and comparisons remain the same.

Reviewer_3BVDOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper aims to perform human motion denoising, along with motion dynamics and anthropometrics estimation using a diffusion framework. The authors use a diffusion approach and augment it with two new pose prior models. The first, DForce, aims to denoise the body dynamics, including the joint torques and ground reaction forces, while the second, DBone, aims to denoise the estimated Anthropometrics.

The framework takes in a noisy observation of the pose sequence, and uses a combination of three losses to iteratively denoise the estimate of the full pose sequence. The first is a data loss that combines the observation's measurement model with a penalty on the residual force at the root joint. The residual force at the root is computed using Lagrangian dynamics.

The second is the proposed pose priors, which aim to denoise the current estimates of force features and anthropometrics. These pose priors were trained using existing datasets. The third is a smoothing loss that penalizes acceleration.

The authors also curate a force-annotated fencing dataset with full motion capture. This is used to compare the baselines using real ground reaction forces captured using a force plate.

The proposed method was tested on a few commonly used datasets in the literature and quantitatively compares on par with other baselines.

Critical strengths

The main strength of this paper is its curated force-annotated mocap video dataset for fencing, which may prove useful for future work.

Critical weaknesses

The main weakness of this paper is the lack of novelty in the proposed approach. Various components of the main highlighted contributions seem to have been done separately in other landmark works. This has led to the results in the Evaluation Section not being substantially better than the baselines. The primary contributions of the Dynamic Pose Priors, namely DBone and DForce, were trained using datasets that were previously used in other works for Human Dynamics Estimation.

The second main weakness is that physics-based pose priors have been proposed earlier in human pose estimation and synthesis, including PhysDiff [1] and Yi et al. [2]. PhysDiff uses Reinforcement Learning to address the lack of training data, while [2] computes them analytically. These approaches can even compute External Reaction Forces for arm and non-foot body contacts, unlike this paper.

[1] Yuan et al. PhysDiff: Physics-Guided Human Motion Diffusion Model. In ICCV 2023.

[2] Yi et al. Improving Global Motion Estimation in Sparse IMU-based Motion Capture with Physics. In SIGGRAPH 2025.

Questions for authors
  1. See Weaknesses

  2. Minor Clarification: In Section 2, the authors claim that the SKEL model was proposed to bridge the gap between the typical SMPL body model and an anatomically accurate physics model. But in Section 3.2, the authors claim that adding scapular anthropometric data to SKEL is required.

1 comment

Dear Reviewer 3BVD. Thank you for taking the time to review our paper. We hope our comments below can help clarify your questions.


(Q1) See Weaknesses.

On lack of novelty.
While sharing similarities, our method is different from the two example works [1,2] you stated that significantly overlap in contribution with ours. Neither of those methods proposes kinematics nor kinetics priors using diffusion models. And our use of diffusion models is different from that in [1] and the way you interpreted our work in your summary.

PhysDiff [1] trains a conditional diffusion model to approximate the posterior distribution of kinematics, from which they sample from. During certain sampling steps, the denoised kinematics output is taken out and "corrected" in a separate physics simulator using an imitation learning policy trained via reinforcement learning, before feeding into the next step.

Our method extends the variational inference framework in [3,4] to inverse dynamics, using diffusion models as unconditional priors for regularization, to maximize evidence/marginal likelihood (lower bound).

Yi et al [2] trains an LSTM-based pose estimator to predict motion from IMUs, then refines the output via physics-optimization. Their proposed "physics" prior is used to constrain global degrees of freedom, not kinetics. Also, being able to compute "more" contact points does not always imply the method is better. For example, our method shows significantly improved feet-ground contact force estimation compared to PhysPT, which models full-body contact. The additional contact points may not be meaningful if their values are inaccurate. The reason for implementing feet-ground contact only is training and evaluation data scarcity. Incorporating additional contact points to human dynamics modeling is straightforward.

On lack of substantial improvements over previous methods. In Sec 4 and A.1, we showed noticeable improvements over previous methods, most notably in terms of kinetics estimation with lower force errors (Tables 2 and 6), which was also not quantitatively evaluated in previous methods.

On the use of existing public datasets. We trained and evaluated our method on existing public datasets to allow for a more standardized comparison with previous works.

[3] Mardani et al, "A variational perspective on solving inverse problems with diffusion models," ICLR, 2024.
[4] Lu et al, "DPoser-X: Diffusion model as robust 3d whole-body human pose prior," CVPR, 2025.


(Q2) Minor Clarification: In Section 2, the authors claim that the SKEL model was proposed to bridge the gap between the typical SMPL body model and an anatomically accurate physics model. But in Section 3.2, the authors claim that adding scapular anthropometric data to SKEL is required.

This is not a typo. SKEL is a kinematic model that includes scapular segments. Hence, in order to model SKEL dynamics, we needed to incorporate mass and inertial properties for the scapulae, which were absent in previous anthropometrics estimation approaches for human models without scapular segments.

Another related correction in your summary. DBone does not denoise anthropometrics, but rather is a pose prior for SKEL's local degrees of freedom.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.