WalkTheDog: Cross-Morphology Motion Alignment via Phase Manifolds

How to read this

Category

Method: unsupervised cross-morphology motion alignment via phase manifolds

Contributions

• Proposes a phase manifold of multiple closed curves (one per latent amplitude) to capture motion periodicity and semantics independent of skeleton
• A vector-quantized periodic autoencoder that learns a shared phase manifold across morphologies (e.g. human and dog) without supervision
• Combines the manifold with an improved motion matching framework for timing/semantics-aligned retrieval, transfer, and stylization

Context

Synthesizes the phase-manifold idea of Starke et al.'s DeepPhase with the cross-character retargeting goals of Aberman et al.'s Skeleton-Aware Networks, replacing a sparse high-dimensional latent with a quantized periodic one shared across morphologies.Builds on: Skeleton-Aware Networks for Deep Motion Retargeting · DeepPhase: Periodic Autoencoders for Learning Motion Phase Manifolds

Correctness

Relies on the assumption that a discrete bottleneck and shallow network force semantically similar motions onto the same curve; demonstrated unsupervised on human and dog data across retrieval, transfer, and stylization, with code released, though alignment quality for highly dissimilar morphologies is the open question.

Clarity

Idea is elegantly motivated but technically involved; a first pass conveys the shared-manifold concept, a second pass is needed for the VQ periodic autoencoder.

How to read it

Focus on the phase-manifold structure and why vector quantization plus a shallow bottleneck drives alignment; a second pass on the autoencoder and motion matching is worthwhile, and code is available to inspect.