Technical Summary: PoseShield: Neural Collision Fields for Human Self-Collision Resolution

Problem Statement

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. While classical collision-handling methods exist in geometry processing and physical simulation, they typically operate directly in mesh space by optimizing raw vertex positions. These approaches are not naturally compatible with learning-based SMPL pipelines where the optimization variable is the pose parameter $\theta$ rather than mesh vertices. Furthermore, classical methods often require a collision-free reference configuration, which is generally unavailable in modern pose-prediction pipelines. Existing learning-based approaches often rely on heuristic penalties or soft constraints that lack the theoretical regularity conditions required for robust, gradient-based constrained optimization.

Methodology

The authors propose PoseShield, a neural collision constraint defined directly in the low-dimensional SMPL pose space. The core of the methodology involves formulating collision resolution as a constrained optimization problem and connecting the learned constraint to the Eikonal equation.

1. Problem Formulation

Given a self-colliding SMPL configuration $(\beta, \theta_0)$, the goal is to find a corrected pose $\theta$ that minimizes the distance to the original pose while satisfying a collision-free constraint:
$$\theta^\star = \arg \min_{\theta} d_{\mathrm{SMPL}}(\theta, \theta_0) \quad \text{subject to} \quad \iota(\mathcal{M}(\beta, \theta)) = +1$$
where $\iota$ is a binary collision indicator. Since $\iota$ is non-differentiable, the authors introduce a surrogate, learnable neural constraint function $g(\theta)$, transforming the problem into:
$$\theta^\star = \arg \min_{\theta} d_{\mathrm{SMPL}}(\theta, \theta_0) \quad \text{subject to} \quad g(\theta) \ge C_l$$

2. Theoretical Foundation: Eikonal Regularization

For gradient-based constrained solvers (e.g., SLSQP, Augmented Lagrangian) to converge reliably, the constraint function must satisfy the Linear Independence Constraint Qualification (LICQ), which requires a non-vanishing gradient near the constraint boundary. The authors establish a connection between this requirement and the Eikonal equation.

They propose training $g(\theta)$ to approximate a Signed Distance Function (SDF) to the collision boundary in pose space. By enforcing an Eikonal regularization term during training:
$$\mathcal{L}_{\text{grad}} = \mathbb{E}_{\theta \sim \Omega} \left[ \left| \|\nabla_{\theta} g(\theta)\| - 1 \right| \right]$$
the method ensures that $\|\nabla g\|$ is bounded away from zero throughout the pose space. Theoretically, this satisfies the LICQ globally, providing convergence guarantees for the optimization process. The authors prove that minimizing this loss bounds the volume of the pose-space regions where LICQ fails.

3. Training and Optimization

• Training Objective: The network is trained using a combination of sign supervision (ensuring $g(\theta) > 0$ for collision-free poses and $g(\theta) < 0$ for colliding ones) and the Eikonal loss. A Temporal Difference (TD) variant of the gradient loss is also introduced to improve training stability.
• Active Learning: To accurately capture the decision boundary, the authors employ an active learning strategy that incrementally augments the training set with samples near the zero-level set of $g(\theta)$.
• Motion Sequence Extension: The learned constraint is extended to human motion sequences. Instead of retraining the underlying motion generator, PoseShield acts as a generator-agnostic post-hoc corrector. It optimizes the motion sequence to satisfy the collision constraint while preserving motion fidelity via a distance metric that includes pose parameters, joint positions, and temporal velocity.

Key Contributions

1. Principled Framework: The paper formulates self-collision resolution in SMPL pose space as a constrained optimization problem with a learnable differentiable constraint, providing theoretical convergence guarantees for gradient-based solvers under specific assumptions.
2. PoseShield: A neural collision constraint trained with Eikonal regularization. The authors prove that this training objective quantitatively bounds the volume of regions where the LICQ fails, linking training accuracy to solver reliability.
3. Generator-Agnostic Correction: The method serves as a post-hoc corrector for human motion sequences without requiring retraining of the underlying motion model, enabling temporally consistent collision resolution.
4. Benchmarking: The authors introduce the Humans with Collisions (HwC) dataset, containing nearly 931k SMPL poses with self-collisions, to facilitate benchmarking.

Experimental Results

Experiments were conducted on the HwC dataset and the PROX dataset, comparing PoseShield against baselines including Torch-mesh-isect, a classifier-based baseline, COAP, and VolumetricSMPL.

• Success Rate (SCC): PoseShield achieved a 95.8% success rate on the HwC dataset, significantly outperforming the best baseline (COAP at 44.6%).
• Penetration Depth Reduction (PDR): The method achieved a PDR of 98.2%, indicating highly effective removal of self-intersections.
• Geometric Fidelity: PoseShield maintained a low Mean Vertex Distance (MVD) of 0.059, demonstrating that it resolves collisions while keeping the corrected pose close to the original input.
• Motion Sequences: In motion correction tasks, PoseShield provided the best balance between collision removal and motion quality, outperforming direct optimization and COAP-based approaches in terms of residual penetration depth and motion artifacts (e.g., foot skating).

Significance and Claims

The paper claims that PoseShield provides a theoretically grounded solution to a persistent problem in human modeling. By connecting collision handling to the Eikonal equation, the authors ensure the numerical stability and robustness of the optimization process, a property often lacking in heuristic penalty-based methods. The ability to operate directly in the low-dimensional pose space, rather than mesh space, allows for seamless integration with learning-based pipelines. Furthermore, the method's generator-agnostic nature allows it to function as a universal refinement module for diverse SMPL-based scenarios, from static pose correction to complex motion synthesis, without modifying the underlying generative models.