Skinned Motion Retargeting with Spatially Adaptive Interaction Guidance

How to read this

Category

Geometry-aware motion retargeting preserving self-contact and near-body interaction across differently proportioned skinned characters

Contributions

• Identifies that prior proximity-matching retargeting methods rely on static predefined anchor correspondences, which fail when the target character has exaggerated or very different body proportions because the corresponding region may be geometrically unreachable
• Introduces an Adaptive Anchor Sampling module that uses a Transformer-based residual predictor plus differentiable soft projection to reposition anchors onto the target mesh at kinematically reachable, semantically plausible locations, conditioned on deformed source anchors that encode the source character's pose-dependent spatial structure
• Feeds the adapted anchors into a Proximity-based Retargeting module, a graph-based autoencoder that predicts target skeletal motion preserving the source's spatial configuration, trained with an alternating scheme that optimizes anchor adaptation and motion retargeting in turn

Context

The work builds directly on proximity-matching retargeting approaches that define fixed anchor regions on the character surface, and on contact-aware retargeting such as CAR, extending both by making the anchor correspondences adaptive instead of static. It sits alongside skeleton-aware and skeleton-agnostic motion-embedding retargeting work that solves the related but distinct problem of transferring joint transforms across differing skeleton hierarchies.

Correctness

The paper reports outperforming prior state of the art in preserving interaction fidelity across diverse character geometries; it is a peer-reviewed ACM SIGGRAPH-track paper, and the archive note's summary of dynamically repositioning anchors onto reachable regions matches the abstract directly. The specific quantitative margins were not visible in the pages read here, so they were not verified.

Clarity

Clearly written for a graphics research audience, and the motivating failure case (anchors landing on unreachable regions of an exaggeratedly proportioned target) is easy to follow even for a working animator. The Transformer-based anchor refinement and graph-autoencoder retargeting architecture need some machine learning background to follow in full.

How to read it

First pass: abstract and Figure 1's teaser showing anchor repositioning across body shapes. Second pass: the introduction for the motivating failure case and related work for how it differs from CAR and static proximity matching. Third pass: the Adaptive Anchor Sampling and Proximity-based Retargeting method sections plus the evaluation for fidelity comparisons.