Skeleton Subspace Skin Penetration Removal

How to read this

Category

Interactive kinematics-based penetration correction for skeleton-subspace-skinned characters (short paper)

Contributions

• Formulates skin penetration removal as a constrained least-squares multi-target inverse kinematics problem on the skeletal pose (SSMIK), adjusting bones rather than moving skin vertices directly, preserving the near-rigid look of SSD models
• Introduces on-the-fly clustering of penetrated vertices by cosine similarity of their deformation-gradient Jacobians, collapsing a dense multi-target IK problem into a lower-dimensional, numerically stable one
• Implements the method as a Maya 2023 C++ plug-in and reports interactive performance (collision, clustering, and pose-optimization steps each under 5ms) across test scenes with up to 26 joints

Context

SSMIK extends classical mesh inverse kinematics and mesh puppetry techniques into a skeleton-subspace-only formulation, standing alongside physics-based contact-response methods like the soft character technique and contact skinning as a purely kinematic alternative that the authors argue avoids their numerical instability under deep penetration.

Correctness

This is a four-page short paper with a modest experimental section: three test models, timing measured on a single machine, no production-scale asset or user study. The clustering similarity threshold (0.8) is set empirically from a supplemental comparison rather than derived analytically, so how well it generalizes beyond the tested scenes is untested.

Clarity

Dense but tightly scoped, typical of the Eurographics short-paper format. The Jacobian pseudoinverse and clustering math is compact and assumes comfort with skeleton-subspace deformation notation, though a working rigger can follow the algorithm steps in Section 3 without difficulty.

How to read it

First pass: abstract, Figure 1's overview of penetration detection and clustering, and the two bullet contributions. Second pass: Sections 3.1 to 3.3 for the penetration formulation and the Jacobian-based pose update. Third pass: Section 3.4's clustering algorithm and Table 1's timings if reimplementing the plug-in.