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

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