← Anthony Roumi

Generalizable Motion Retargeting Using Deep Skeletal Networks

Imitation Learning · Motion Representation

Code


Motion retargeted from group B to group A

Motion from a group B skeleton retargeted onto a group A skeleton.

Motivation

Retargeted motion data across arbitrary pairs of skeletal domains, without requiring a domain-specific mapping to be hand-authored for every new source/target skeleton pair.

Method

A deep skeletal network learns a shared motion representation that generalizes across differing joint topologies and proportions, allowing motion captured on one skeleton (e.g. a human actor) to be retargeted onto another (e.g. a simulated humanoid) while preserving the semantics and style of the original motion.

Results

Source motion

Retargeted motion