PaperlayerView exact revision on arXiv

Source-linked AI summary

PoseShield: Neural Collision Fields for Human Self-Collision Resolution

Zhengyuan Li, Zeyun Deng, Yifan Shen, Liangyan Gui, Miaolan Xie, Joseph Campbell, Xifeng Gao, Kui Wu, Zherong Pan, Aniket Bera

arXiv:2606.29686v2cs.CV

TL;DR

Self-collisions in SMPL-based reconstruction and motion synthesis reduce physical plausibility and motivate reliable post-hoc correction. PoseShield learns an Eikonal-regularized neural collision constraint in pose space and uses constrained optimization, substantially outperforming prior post-hoc baselines in collision-resolution success.

  • Problem

    SMPL-based reconstruction and motion synthesis can produce self-intersections, motivating reliable post-hoc correction decoupled from specific generative priors.

  • Method

    PoseShield learns a differentiable collision constraint directly in SMPL pose space and uses Eikonal regularization with gradient-based constrained optimization.

  • Results

    PoseShield substantially outperforms prior post-hoc collision-handling baselines on the HwC and PROX datasets.

  • Takeaways & Limitations

    The learned constraint provides a generator-agnostic post-hoc corrector for human motion sequences without retraining the underlying motion model.

  • Takeaways & Limitations

    The formulation assumes fixed body shape and optimizes only pose, ignoring global translation and rotation because self-collision is invariant to them.

Abstract

from arXiv · show

Self-collision remains a persistent challenge in SMPL-based human pose estimation and motion generation. Under extreme articulations or stochastic motion synthesis, generated meshes frequently exhibit self-penetrations, leading to physically implausible results. We propose PoseShield, a neural collision constraint defined directly in SMPL pose space. We formulate collision correction as a constrained optimization problem and connect the learned constraint with the Eikonal equation. Enforcing Eikonal regularization ensures non-vanishing gradients near the collision boundary, improving numerical stability and robustness of the optimization process. Unlike prior methods that operate in the mesh space or rely on heuristic penalties, our approach operates directly in the low-dimensional space of human poses and is theoretically grounded. The same learned constraint extends to human motion sequences, providing a generator-agnostic post-hoc collision corrector without retraining the underlying motion model. Experiments on a newly constructed SMPL pose benchmark show that our method achieves a 95.8% success rate and outperforms state-of-the-art baselines.

1 Introduction

Self-collision persists across SMPL-based pose reconstruction and motion synthesis, motivating a universal post-hoc correction method decoupled from specific generative priors. PoseShield addresses this need with a differentiable collision constraint learned directly in pose space, theoretically linked to solver reliability and reusable for motion correction.

  • Motivation: Self-collisions affect SMPL-based reconstruction and stochastic motion synthesis, producing body intersections that degrade generated results.COAP identifies intersections in PROX poses, while recent work finds non-negligible intersections in synthesized motions.
  • Motivation: A reliable post-hoc resolver should act as a universal refinement module, independent of generative priors, prompts, or reference images.The goal is geometric consistency across diverse SMPL-based scenarios.
  • Method: PoseShield formulates correction as nearest collision-free pose search using a differentiable neural collision constraint defined directly in SMPL pose space.The constrained problem can be solved with gradient-based methods such as SLSQP and augmented Lagrangian algorithms.
  • Theory and training: Eikonal regularization trains PoseShield to satisfy solver-relevant assumptions, including reliable collision signs and nonvanishing gradients required by LICQ.The Eikonal loss bounds the volume of pose-space regions where LICQ fails, connecting training accuracy with solver reliability.
  • Applications and evaluation: The learned constraint extends to temporally consistent motion correction without retraining the underlying generator and substantially outperforms prior post-hoc baselines on HwC and PROX.This provides a generator-agnostic post-processing module for human motion synthesis.

2 Related Work

Prior SMPL self-collision methods progressed from coarse geometric proxies to mesh-level penetration penalties, while general collision-resolution techniques optimize raw mesh vertices rather than pose variables. PoseShield is positioned against these approaches by modeling self-collision as a neural field for constrained pose-space optimization, drawing on neural Eikonal/SDF methods.

  • Human Self-Collision in SMPL-based Modeling: SMPL self-collision research began with coarse geometric heuristics, including capsule-based primitive proxies, for simplifying intersection queries.These methods target issues arising in extreme articulations and depth ambiguities in monocular reconstruction.
  • Human Self-Collision in SMPL-based Modeling: Later methods introduced mesh-level penetration penalties, including SMPLify-X, distance-field-style losses, accelerated intersection detection, PROX scene-aware constraints, and sphere-based motion-generation proxies.The sphere-based approach incorporates self-intersection losses during training.
  • General Mesh Collision Resolution: Classical collision-resolution methods use penalty energies, interior-point formulations, or Repulsive Shells while optimizing raw mesh vertices directly.The passage identifies this mesh-space operation as a limitation for pose-space optimization and notes that such methods often require a collision-free reference configuration.
  • Neural Solution of the Eikonal Equation: Neural Eikonal solvers with sign constraints compute signed distance fields, enabling representation of complex geometry in modern 3D generative models.These methods generally assume a low-dimensional 3D Euclidean domain, whereas PoseShield designs a field for the specific constraint of self-collision in pose-space optimization.

3 Problem: SMPL Self-Collision Resolution

This section defines SMPL self-collision resolution as finding a collision-free pose for a fixed body shape while minimizing deviation from an original colliding pose. SMPL meshes are generated from shape and pose parameters, and collisions are identified after mesh decoding.

  • SMPL formulation: An SMPL mesh is generated as X = M(β, θ) from shape parameters β and pose parameters θ, with fixed mesh connectivity.The mesh connectivity T is predefined and independent of β and θ.
  • Collision setting: Generated SMPL meshes may self-collide because of imperfect motion capture or errors from neural motion predictors.In practical motion-correction scenarios, body shape commonly remains fixed across frames.
  • Collision setting: A classical collision detector produces a binary indicator that equals −1 for self-penetration and +1 otherwise.The pose is decoded into a mesh before collision detection is applied.
  • Resolution objective: The correction seeks a pose whose decoded mesh is collision-free while remaining visually close to the original configuration.The discrepancy is measured by dSMPL between poses of a fixed shape.
  • Resolution objective: The problem minimizes dSMPL(θ, θ0) subject to the decoded mesh having a positive collision indicator.Formally, θ⋆ = arg minθ dSMPL(θ, θ0) subject to ι(M(β, θ)) = +1.

4 Learning a Differentiable Collision Constraint

PoseShield replaces the non-differentiable binary collision indicator with a learnable differentiable constraint in SMPL pose space, enabling gradient-based collision-resolution optimization. Smoothness, feasibility consistency, and approximate Eikonal behavior support convergence guarantees, while Eikonal regularization quantitatively bounds regions where LICQ may fail.

  • Motivation: The binary collision indicator is discontinuous with zero gradient almost everywhere, preventing standard gradient-based constrained solvers from being applied.Such solvers require at least C1-smooth constraints and constraint qualifications for reliable convergence to KKT points.
  • Surrogate Constraint: PoseShield learns g(θ), whose superlevel set approximates collision-free poses, and minimizes SMPL pose distance subject to g(θ) ≥ C_l.C_l = 0 is theoretically optimal for a perfect constraint; adjusting it trades geometric fidelity against collision-resolution effectiveness.
  • Convergence Analysis: Under smoothness, feasibility consistency, and approximate Eikonal assumptions, gradient-based solvers satisfy global LICQ and converge from any starting pose to a first-order KKT point.The stated worst-case first-order iteration complexity is O(ε−2), matching unconstrained smooth optimization.
  • Convergence Analysis: The approximate Eikonal condition requires ∥∇θg(θ)∥ to remain near 1, ensuring nonzero gradients and global LICQ under the stated assumptions.With κ < 2, exact-Hessian SQP converges quadratically locally, while a suitable BFGS variant converges superlinearly.
  • Eikonal Regularization: If the expected Eikonal loss satisfies L_grad ≤ ε, the failure-region probability obeys P(θ ∈ S_δ) ≤ ε/δ.This links training accuracy to the pose-space volume where LICQ is guaranteed by the approximate Eikonal property.

5 Evaluation

PoseShield substantially improves static-pose collision resolution over the evaluated baselines, achieving higher success and penetration-depth reduction while maintaining smaller pose deviations. Ablations support its weighted pose distance, Eikonal-related loss design, controllable collision–fidelity trade-off, and extension to motion sequences.

  • 5.1 Static Pose Collision Resolution: PoseShield consistently removes self-collisions in qualitative examples, whereas Torch-mesh-isect fails in all three cases and COAP leaves minor residual intersections in one case.COAP nearly resolves the first case but does not fully eliminate the remaining intersections.
  • 5.1 Static Pose Collision Resolution: 0.446 to 0.958: PoseShield increases success rate on the HwC dataset while achieving significantly lower MVD than the compared method.The result indicates collision resolution with smaller pose deviations and empirically approximates minimal-distance corrections.
  • 5.2 Ablation and Properties: LTD alone achieves the best overall performance across metrics, while omitting both loss terms yields poor performance because the learned function fails to approximate the Eikonal equation.The authors conjecture that Lgrad introduces second-order derivatives, increasing training instability and hindering convergence.
  • 5.2 Ablation and Properties: PoseShield exposes a controllable trade-off: increasing the constraint margin raises SCC but also increases MVD, while the learned constraint generalizes to motion sequences without noticeable artifacts.The motion-sequence correction preserves overall motion and avoids noticeable artifacts.

6 Conclusion

PoseShield is presented as a neural collision constraint in SMPL pose space for post-hoc self-collision resolution. Its Eikonal-based formulation theoretically grounds constraint learning and supports feasible, numerically stable constrained optimization.

  • PoseShield defines a neural collision constraint directly in SMPL pose space for post-hoc self-collision resolution.
  • The method connects collision handling with the Eikonal equation, providing theoretical grounding for neural constraint learning.
  • Eikonal-regularized constraint functions satisfy the LICQ, ensuring feasibility and numerical stability in pose-space constrained optimization.

A Humans with Collisions Dataset

The HwC Dataset is constructed in SMPL pose space by perturbing natural MotionFix poses to produce self-intersecting samples, while distinguishing perceptually relevant collisions from LBS artifacts. It uses train/test splits and analyzes collision severity through penetration depth, with a 500-mesh evaluation subset sampled for computational efficiency.

  • Collision labeling: Collision labels distinguish perceptually relevant interactions across distinct body parts from underarm and behind-knee intersections caused primarily by LBS artifacts.Examples of relevant interactions include hand–body, hand–leg, and leg-related contacts.
  • Dataset construction: The dataset augments MotionFix meshes with Gaussian noise and Gram–Schmidt orthonormalization to generate self-intersecting poses in SMPL pose space.Global translation and rotation are excluded, and remaining joint rotations use a 6D representation.
  • Dataset construction: The generated poses are split into training and test sets with a 9:1 ratio.The split is designed to keep colliding poses close to the natural distribution of valid human poses.
  • Dataset analysis: Penetration-depth statistics cover a wide range of collision severities and exclude non-colliding poses.The distribution is reported only for self-colliding samples.
  • Evaluation protocol: 500 self-penetrating meshes are randomly sampled for baseline evaluation because testing the full dataset is computationally expensive.This follows the practice of previous work.

B Theoretical Analysis … B.3 Assumptions on the Neural Collision Field

The theoretical analysis formulates SMPL self-collision correction as constrained pose optimization, derives an indicator-induced pose-space SDF, and establishes assumptions for a smooth neural collision-field surrogate. The surrogate uses feasibility consistency and an approximate Eikonal property to support collision resolution in bounded 6D pose space.

  • B.1 Problem Formulation: The analysis formulates collision resolution as finding the closest collision-free pose to a colliding SMPL configuration under a fixed shape.The optimized variable is the pose θ, while global translation and rotation are ignored because they do not affect self-collisions.
  • B.1 Problem Formulation: The pose domain is the intersection of each joint’s non-degenerate 6D rotation domain and a bounded box Ω_B = [-B,B]^(J×6).Degenerate inputs outside the valid 6D domain are treated as infeasible by the extended collision indicator.
  • B.1 Problem Formulation: The exact collision indicator assigns −1 to colliding or degenerate poses and +1 to non-degenerate collision-free poses, defining the feasible set F_β.For valid poses, the indicator is obtained by decoding the SMPL mesh and applying an exact self-intersection test such as FCL.
  • B.2 Indicator-Induced Signed Distance Function: The collision indicator induces a pose-space signed distance function whose sign separates collision-free and infeasible regions.The function is defined as distance to the infeasible set minus distance to the feasible set, and is well-defined on Ω_B.
  • B.2 Indicator-Induced Signed Distance Function: At almost-everywhere differentiability points away from the zero level set, the indicator-induced SDF satisfies the Eikonal property ||∇ϕ_β(θ)||_2 = 1.The function is Lipschitz continuous and differentiable almost everywhere.
  • B.2 Indicator-Induced Signed Distance Function: Under Euclidean pose distance, the optimal correction value equals the distance from an infeasible pose to the closure of the feasible set.A minimizer in the feasible set exists when F_β is closed.
  • B.3 Assumptions on the Neural Collision Field: The approximate Eikonal condition makes g an approximate pose-space SDF with non-vanishing boundary gradients, yielding the surrogate constraint g(θ) ≥ 0.The resulting optimization minimizes pose discrepancy subject to the learned field’s nonnegative superlevel set.

B.4 Convergence Analysis · B.5 Theoretical Justification of the Eikonal Loss

Under the stated assumptions, the constrained SMPL pose problem has global and local SQP convergence guarantees, with complexity matching unconstrained smooth optimization. The Eikonal loss is theoretically justified because its small expected gradient-norm error bounds the volume of regions violating the approximate Eikonal condition.

  • B.4 Convergence Analysis: O(ε^-2) is the worst-case iteration complexity for obtaining an ε-approximate KKT point, matching unconstrained smooth optimization.The approximate Eikonal condition uniformly bounds the constraint Jacobian’s minimum singular value away from zero, preventing conditioning-related degradation.
  • B.4 Convergence Analysis: The approximate Eikonal condition removes infeasible stationary-point stagnation by ensuring every infeasible point admits a feasible direction in the linearized constraint.The direction d = t∇g(θ), with t ≥ |g(θ)|/(1-δ)^2, makes the linearized constraint feasible and the SQP subproblem strictly feasible.
  • B.4 Convergence Analysis: For an infeasible initial pose, the local KKT multiplier is strictly positive and bounded by 0 < 2∥θ⋆−θ0∥/(1+δ) ≤ λ⋆ ≤ 2∥θ⋆−θ0∥/(1−δ).This establishes strict complementarity at the local minimizer.
  • B.4 Convergence Analysis: When κ < 2, the full Lagrangian Hessian is positive definite, yielding second-order sufficient conditions and quadratic local convergence for exact-Hessian SQP.A BFGS variant satisfying the Dennis–Moré condition converges superlinearly.
  • B.5 Theoretical Justification of the Eikonal Loss: The Eikonal regularization Lgrad is defined as the expected absolute deviation |∥∇g(θ)∥−1| under a uniform measure over ΩB.The analysis uses this expected error to quantify where the approximate Eikonal condition fails.
  • B.5 Theoretical Justification of the Eikonal Loss: P(θ ∈ Sδ) ≤ ε/δ whenever Lgrad ≤ ε, so minimizing the Eikonal loss bounds the probability volume of regions failing the margin-δ condition.The bound follows directly from Markov’s inequality applied to X(θ) = |∥∇g(θ)∥−1|.

C The Gap between Theory and Practice

The convergence guarantees rely on assumptions that are only approximately satisfied in practice. Feasibility accuracy reaches 93.9%, while approximate Eikonal behavior holds for 95% of test poses, and sampling and preprocessing introduce further domain mismatches.

  • Assumption discrepancies: Theoretical convergence guarantees depend on idealized assumptions that are only approximately satisfied in practice, although activation choices show no significant performance difference.Softplus yields a C∞ network, whereas ReLU and ELU are smooth almost everywhere.
  • Assumption discrepancies: 93.9% test-set accuracy is achieved for collision indication, revealing failure of Feasibility Consistency.This corresponds to the practical failure of Assumption 5.
  • Assumption discrepancies: 95% of test samples satisfy the approximate Eikonal property with δ = 0.1, despite exact satisfaction not being guaranteed.Minimizing Lgrad bounds the volume of regions where Assumption 6 fails; Fig. 8 evaluates approximately 92k samples after removing 0.5% outliers on each side.
  • Sampling: Training samples follow a data-induced distribution from noisy motion data and projected valid 6D rotations rather than the uniform distribution assumed theoretically.This concentrates samples in practically relevant regions but does not exactly match the uniform measure over ΩB.
  • Network input preprocessing: Gram–Schmidt orthonormalization keeps optimization inputs in-distribution but restricts the effective domain to SO(3)J and may limit approximation capacity over ΩB.The resulting function is constant on level sets of π6D, satisfying ˜g(θ1) = ˜g(θ2) whenever π6D(θ1) = π6D(θ2).

D Limitations and Future Work · E Our Model as a Classifier

The paper identifies semantic fidelity as a limitation of geometric pose and motion distances and proposes integrating semantic metrics in future work. It also shows that the learned collision constraint can function as a classifier using the sign of g, evaluated by accuracy and false negative rate.

  • D Limitations and Future Work: Geometric metrics alone measure distances between poses and motions, although practical users may prioritize semantic fidelity.The paper gives hand-to-head contact as an example where semantic fidelity may matter.
  • D Limitations and Future Work: Integrating semantic distance metrics is proposed as a valuable direction for future work.
  • D Limitations and Future Work: The method can be seamlessly extended to parametric human models beyond SMPL, including Momentum Human R.
  • E Our Model as a Classifier: The sign of g indicates a sample’s collision status, enabling the method to operate as a collision detector.
  • E Our Model as a Classifier: Classifier evaluation uses prediction accuracy, which measures whether collision labels are correctly predicted.
  • E Our Model as a Classifier: Classifier evaluation also uses false negative rate, measuring colliding meshes predicted as collision-free.
  • E Our Model as a Classifier: The results show that the method can serve as a classifier with performance comparable to a standard binary classifier.These results are reported in Table 2.

F Human Motion Collision Resolution: Implementation Details · G Details of Baseline Implementation · H Details of Active Learning

The implementation preserves source-motion fidelity through pose-space, joint-space, and temporal losses, while baseline comparisons use specified model mappings and stopping rules. Active learning targets the collision decision boundary by collecting optimizer iterates near the learned zero-level set.

  • F Human Motion Collision Resolution: Implementation Details: Motion fidelity is maintained with a pose-space distance between optimized and source sequences, averaged over T+1 frames.The pose loss uses the mean squared Euclidean distance between corresponding pose parameters.
  • F Human Motion Collision Resolution: Implementation Details: Joint-space and temporal losses compare SMPL joint positions and their frame-to-frame changes against the source motion.SMPL forward kinematics converts each pose into 3D joints before computing positional and velocity discrepancies.
  • F Human Motion Collision Resolution: Implementation Details: The motion distance combines feature, joint-position, and velocity losses to preserve pose similarity, 3D fidelity, and temporal consistency.The default weights are λjoint = 1 and λvel = 0.1.
  • G Details of Baseline Implementation: VolumetricSMPL follows its official implementation and evaluates mapped SMPL test rotations under SMPL-X, excluding 140 out of 500 HwC samples.Excluded samples collide in SMPL but not in SMPL-X.
  • G Details of Baseline Implementation: COAP optimizes body pose with SGD using self-penetration loss plus a pose prior, stopping at its threshold or after 200 iterations.Torch-mesh-isect instead receives a 3-minute maximum runtime because its original implementation lacks internal stopping.
  • G Details of Baseline Implementation: The classifier baseline trains on HwC and constrains pose optimization to classifier outputs greater than 0.5 while minimizing distance from the initial pose.It uses the classifier as a surrogate feasibility test in the constrained solver.
  • H Details of Active Learning: Random pose augmentation can be distribution-biased and undersample the near-boundary region where g(θ) must match the exact collision indicator.Accuracy far from the zero-level set is less critical because those regions are mainly traversed during intermediate optimization steps.
  • H Details of Active Learning: Active learning collects intermediate optimizer iterates and converged solutions to concentrate samples near g(θ)=0 and improve the learned decision boundary over iterations.The pipeline repeatedly samples poses, solves the auxiliary optimization, and augments training data with the resulting iterates.

I More Human Motion Collision Resolution Results

On 100 highly penetrating MotionFix sequences, PoseShield achieves the best balance between collision removal and motion quality. Compared with direct optimization and COAP (DNO), it preserves more natural motion during collision resolution.

  • Data: 100 MotionFix sequences with the largest total penetration depths form the evaluation set.The sequences are selected for their severe self-penetration.
  • Metrics: The evaluation measures jitter, Foot Skating Ratio (FSR), Residual Penetration Depth (RPD), and Motion Feature Distance (MFD).FSR captures foot-skating artifacts, RPD measures residual interpenetration severity, and MFD measures feature-space discrepancy from the original motion.
  • Results: Direct optimization is fastest and closest to the input motion in feature space, but leaves more residual collisions and increases jitter and foot skating.It optimizes the motion sequence itself rather than the diffusion model’s input noise and adds a smoothness term with λsmooth set to 0.5.
  • Results: COAP (DNO) improves smoothness over direct optimization.It uses the same DNO optimization algorithm while replacing the collision term with COAP self-collision loss, sampling 50 points per body part.
  • Results: PoseShield provides the best balance between collision removal and motion quality.The quantitative comparison is reported in Table 3.
Loading 2606.29686v2…