Unified Cross-Structural Motion Retargeting for Humanoid Characters

Area of Science:

• Computer Graphics
• Animation
• Machine Learning

Background:

• Current motion retargeting methods struggle with varying skeletal structures.
• Existing techniques often require structure-specific model designs and training.
• A need exists for a unified approach to handle diverse character skeletons.

Purpose of the Study:

• To develop a unified dynamic graph network for motion retargeting across unseen skeletal structures.
• To enable efficient and unsupervised cross-structural motion transfer.
• To handle intricate whole-body movements, including hand animations.

Main Methods:

• Utilized a dynamic graph transformation module to transfer latent motion features between different skeletal structures.
• Modeled torso and hand joints as graphs in a unified manner for comprehensive motion retargeting.
• Employed an unsupervised learning strategy with unpaired motion data from diverse structures.

Main Results:

• The proposed unified dynamic graph network demonstrated superior performance compared to existing methods.
• Achieved significant improvements in data efficiency for training motion retargeting models.
• Successfully performed cross-structural motion retargeting on both known and previously unseen skeletal structures.

Conclusions:

• The unified dynamic graph network offers a flexible and effective solution for motion retargeting across diverse skeletal structures.
• The method advances animation production and virtual reality by enabling seamless motion transfer.
• Unsupervised learning with unpaired data significantly enhances the practicality and applicability of motion retargeting techniques.