- 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
-
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.
-
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
-
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 priorablation does not isolate DForce, leaving the additional value of the force diffusion prior unclear. -
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
-
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?
-
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
(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 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.
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.