How Noisy Poses Break Inverse Dynamics: Analysis and Mitigation for Video-Based Joint Torque Estimation

Abstract: Recent advances in monocular 3D human pose estimation enable accurate body tracking from video. However, translating these kinematic estimates into physical quantities, such as joint torques, remains challenging due to noise amplification through inverse dynamics. In this work, we provide a systematic analysis of how pose estimation noise propagates through the inverse dynamics pipeline. We present three key findings: (1) pose noise is amplified by approximately 1,000x when computing joint torques via numerical differentiation, (2) proximal joints (spine, hips) are up to 10x more sensitive to noise than distal joints (wrists, hands), and (3) low-pass filtering before differentiation substantially reduces this amplification. To enable this analysis, we develop SMPL-Dynamics, a fully differentiable inverse dynamics module for the SMPL body model that requires no external physics simulators. Our module supports end-to-end gradient computation, and we demonstrate this through differentiable pose refinement, which reduces torque error by 93% with negligible change in pose.

• The paper introduces SMPL-Dynamics, a fully differentiable Recursive Newton-Euler implementation for SMPL models that propagates gradients from estimated torques back to pose parameters without external simulators.
• The paper finds that realistic pose noise can amplify torque errors by about 1,000×, producing roughly 60 Nm error at 0.05 radians and making proximal trunk joints up to 10× more influential than distal joints.
• The paper shows that 6 Hz Butterworth filtering cuts torque error by about 55%, while differentiable pose refinement reduces error from 59.7 Nm to 4.7 Nm with less than 1% change in pose error.

Overview

This paper addresses a gap between two research communities: computer vision, which treats human motion as a kinematics problem, and biomechanics/physics-based animation, which requires dynamics quantities such as joint torques and ground reaction forces. The authors provide the first systematic quantitative analysis of how noise in video-based 3D pose estimates propagates through inverse dynamics on the SMPL body model (2605.24776). Their central instrument is SMPL-Dynamics, a fully differentiable implementation of the Recursive Newton-Euler Algorithm (RNEA) operating directly on SMPL kinematic trees with de Leva anthropometric body segment parameters, requiring no external simulators such as OpenSim or MuJoCo.

The SMPL-Dynamics module

The pipeline takes SMPL parameters (root orientation, 23 axis-angle joint rotations, root translation) through three stages: (i) assignment of segment masses, centers of mass, and inertia tensors from de Leva tables; (ii) numerical differentiation of joint positions via central finite differences, with angular velocities and accelerations derived from consecutive rotation matrices via the matrix logarithm; and (iii) RNEA traversal from leaves to root accumulating Newton-Euler force and torque equations. Because every stage is implemented in differentiable operations, gradients flow from torque outputs back to pose inputs — a property the authors exploit for optimization-based refinement that non-differentiable tools like OpenSim cannot support directly.

Quantifying noise amplification

The core empirical finding is stark: injecting isotropic Gaussian noise ($\sigma \in [0, 0.2]$ rad) into axis-angle poses of clean walking sequences yields an approximately $1{,}000\times$ amplification factor from pose error to torque error, driven by the $1/\Delta t^2$ scaling of second-order finite differences. At $\sigma = 0.05$ rad — typical of modern estimators such as WHAM and HMR2.0 — mean torque error reaches roughly 60 Nm, which exceeds the maximum hip torque during normal walking (~30 Nm). In other words, at realistic noise levels the noise-induced error dominates the physical signal itself. This quantitatively explains why prior physics-based pose estimation methods (PhysCap, PhysPT, OSDCap) found raw video kinematics inadequate for downstream physics tasks.

Per-joint sensitivity hierarchy

Injecting noise into individual joints reveals a sensitivity ordering that tracks subtree mass accumulation:

Joint	Torque error from isolated noise
spine1	34.4 Nm
spine2	26.0 Nm
spine3	16.5 Nm
hips	~8.6 Nm
distal joints (wrists, hands)	<1 Nm

Proximal joints are up to $10\times$ more sensitive than distal ones because rotation errors at the trunk propagate to the entire upper-body mass distribution. The practical implication is direct: current pose estimators are benchmarked largely on limb accuracy, but for physics applications trunk accuracy should be prioritized, since improving spine estimation yields the highest return for torque quality.

Low-pass filtering as mitigation

Because second-order differentiation amplifies a frequency component at $f$ Hz by $f^2$, broadband estimation noise is amplified far more than low-frequency motion content. Applying a zero-phase 4th-order Butterworth filter before differentiation exploits this spectral separation. The cutoff involves a trade-off: very low cutoffs (2–4 Hz) remove nearly all noise but distort fast movements whose harmonics extend to 6–8 Hz; the authors recommend 6 Hz as a general-purpose default. Under a realistic WHAM-like noise profile (joint-dependent magnitudes, $2\times$ depth-axis amplification reflecting monocular depth ambiguity, and temporal jitter), realistic noise produces higher torque error (95.9 Nm) than uniform noise of equivalent average magnitude (88.4 Nm), since it concentrates error on high-sensitivity joints. Filtering reduces both to ~42 Nm — a 55% reduction — but the residual is dominated by low-frequency pose bias that filtering cannot remove.

Differentiable pose refinement

To address the residual low-frequency bias, the authors optimize noisy poses against a physics loss combining three terms: torque smoothness (penalizing $\|\ddot{\boldsymbol{\tau}}\|$), a physiological magnitude limit via ReLU($\|\boldsymbol{\tau}\| - \tau_{\max}$) with $1{,}000\times$0 Nm, and a regularizer keeping refined poses near the noisy input. After 200 Adam iterations, torque error drops from 59.7 Nm to 4.7 Nm — a 92% reduction (93% is cited in the abstract) — while pose error changes by less than 1% (0.0795 → 0.0787 rad). This result indicates that physically plausible poses lie extremely close to noisy kinematic estimates, and that gradient-based search through a differentiable inverse dynamics module can recover them efficiently without altering the kinematic solution appreciably.

Limitations and open questions

Several caveats bound these results. The analysis is conducted primarily on synthetic Gaussian and simulated WHAM-like noise injected into clean SMPL walking sequences; validation on real video-derived poses with genuine estimator errors, other locomotion modes, and contact-rich activities remains unaddressed. BSPs are fixed from de Leva population tables rather than subject-specific, so model calibration uncertainty — the complementary error source analyzed by Uchida and Seth — is not treated here. The recommended 6 Hz cutoff is activity-dependent, and the refinement experiments use hand-tuned loss weights ($1{,}000\times$1, $1{,}000\times$2, $1{,}000\times$3) whose robustness across motions is not established. Finally, while the conclusion proposes integrating the module as a physics-informed loss inside pose estimator training pipelines, no such end-to-end training experiment is reported.

Conclusion

The paper delivers a controlled, quantitative account of why video-to-torque pipelines fail: quadratic derivative amplification produces a ~$1{,}000\times$4 pose-to-torque error factor, with proximal joints bearing the brunt due to subtree mass accumulation. It shows that Butterworth pre-filtering recovers roughly half the damage and that differentiable RNEA-based pose refinement removes most of the remainder with negligible kinematic change. The main open question left by the work is whether these gains transfer from synthetic noise injection to real monocular pose estimators trained end-to-end with the proposed differentiable dynamics loss.