acceptodds
Under review as a conference paper at ICLR 2027

Beyond Recovery: Evaluating Dynamics and Preservation in Human Motion Post-Processing

Abstract

Motion denoising and post-processing are widely used to refine recovered or generated human motion. Yet evaluations based on positional and acceleration errors can assign similar scores to outputs with different frequency content, local jerk patterns, or changes to already-clean input. We propose a dynamics-aware evaluation protocol that pairs recovery from corruption with preservation of clean motion. Alongside positional error, it adopts two established quantities: log-spectral distance (LSD) between joint-velocity power spectra, and jitter ratio, which compares jerk magnitude with a reference. An identity setting passes clean ground-truth trajectories through each post-processor using the same configuration as its recovery test. We evaluate classical filters and learned post-processors under Gaussian corruption, synthetic video-like corruption, and real estimator errors across multiple motion datasets. Methods with similar Mean Per Joint Position Error (MPJPE) differ substantially in spectral fidelity and clean-input preservation. Spectral and window-level analyses locate the affected frequencies, joints, and time intervals. Re-tuning the classical filters for recovery also increases the changes they make to clean motion. These results show why recovery gains should be evaluated alongside their cost to motion preservation

Primary area:
datasets and benchmarks
Keywords:
Human motion post-processing, Motion denoising, Motion preservation, Evaluation protocols.

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_FgdqOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a dynamic fidelity evaluation protocol for human motion denoising and post-processing. The authors believe that traditional MPJPE and acceleration errors are insufficient to fully reflect motion quality, especially when it is difficult to distinguish between real motion details and noise that are simultaneously suppressed.

Critical strengths

I think the starting point of the entire paper is a good one. The exploration of movement quality assessment is of practical significance. The author not only evaluates whether the algorithm can repair noisy human movements, but also evaluates whether the algorithm destroys the original correct movements through Identity testing.

Critical weaknesses
  1. Why is there no visualization in the entire article? This work specializes in human motion evaluation and can fully demonstrate the comparison between original estimation, different post-processing methods, and Ground Truth. The author can fully display the three-dimensional human skeleton motion, joint trajectories, key time periods, and the corresponding MPJPE, LSD, and JRD at the same time. Moreover, I have not seen any supplementary materials for the "Appendix A/B, Appendix B.9, B.10" mentioned in the paper. The paper was clearly not ready for publication.
  2. As far as I know, LSD, jerk, and spectral analysis all have a long history of research. The main contribution of this paper is to combine these existing tools with clean input identity testing into a new evaluation protocol. However, I am not convinced by the fact that the paper has not fully demonstrated that this protocol can reflect the true quality of human movement better than existing evaluation metrics.
Questions for authors

When the MPJPE of two post-processing results are almost the same, can your LSD/JRD stably determine which human action is more visually natural and kinematically reasonable? Please demonstrate with actual human motion sequences or independent perceptual evaluation.

Reviewer_FwSqOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

Motion post-processing is common strategy to improve the 3D character motion quality obtained from various resource. This paper propose a evaluation protocol to evaluate those motion post processing methods. It suggest use MPJPE, log-spectral distance (LSD) and Jitter Ratio for evaluation metrics. In addition to evaluating the motion recover from corrupted motions, it also suggest use the same metrics to evaluate the motion from passing the ground truth clean motion to the target post processing method. It provides discussion on these metrics and on several benchmarks with different source.

Critical strengths
  1. How to evaluate motion refinement methods is an interesting problem. The propose protocol make some sense to focus on different aspects, not limited to position reconstruction.
  2. It performs a suite of evaluation on several dataset from various origin.
Critical weaknesses
  1. The use of addition metrics for dynamic feature is straightforward. So is the design of a preservation test. Adopting these existing metrics and simple evaluation protocol seems to be trivial. More importantly, the paper does not present a better evaluation. Showing the different metric ranking does not demonstrate the proposes protocol is better.
  2. The paper is built on a assumption that prior methods on motion refinement using only MPJPE as the evaluation metrics, which is no true. jerk metrics and plenty of dynamic or artifacts detection metrics are widely used.
  3. It's hard to tell how the propose evaluation protocol can guide the construction or selection of better motion refinement method.
  4. There's no motion reference example or human perceptual score alignment experiments presented in the paper. It's hard to calibrate what the value actually reveals from the motion.
  5. In table 2, learned post processors introduce severe MPJPE degradation, which seems to be an implementation error.
Questions for authors
  1. Which is the best post processor after all the benchmark evaluation?
Reviewer_7uQ8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper examines the evaluation of human motion post-processing, focusing on whether methods preserve motion characteristics while improving corrupted inputs. It combines existing dynamics measures with tests on noisy and clean inputs to study the trade-offs between recovery and preservation. The contribution is primarily an extension of the evaluation perspective, supported by empirical analyses illustrating differences that conventional error metrics may overlook.

Critical strengths
  1. The paper addresses a relevant evaluation concern: improvements on noisy motion may come at the cost of changing already-accurate inputs. Evaluating recovery and preservation together provides a useful perspective on post-processing behavior.
  2. The empirical comparisons cover both classical and learned methods under different input conditions. These comparisons help illustrate differences in motion characteristics that conventional error metrics may overlook.
Critical weaknesses
  1. The paper provides insufficient evidence for the practical value of its dynamics metrics. The results do not clearly show that LSD/JRD provide useful information about physical or downstream outcomes beyond conventional errors.
  2. The evaluation would benefit from including additional post-processing methods specifically designed to recover motion dynamics, and testing whether the proposed metrics reliably reflect their independently verified quality improvements.
Questions for authors

I am curious about the motivation behind the specific metric definitions. For example, why do the authors formulate JR/JRD based on ratios of average jerk magnitudes rather than directly measuring framewise jerk error? Since similar average magnitudes can conceal differences at corresponding time steps, could the authors clarify the rationale behind this formulation?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.