Dynamic Motion Transition: A Hybrid Data-Driven and Model-Driven Method for Human Pose Transitions.

Area of Science:

• Computer Graphics and Animation
• Artificial Intelligence
• Biomechanics

Background:

• Generating natural human pose transitions is crucial for 3D graphics and animation.
• Existing model-driven methods rely on difficult-to-obtain heuristic knowledge.
• Pure data-driven methods struggle with unseen data and limited generative power.

Purpose of the Study:

• To propose a hybrid approach for robust and accurate virtual human pose transitions.
• To address limitations of existing pure data-driven and model-driven methods.
• To introduce a novel metric for evaluating generative power in motion synthesis.

Main Methods:

• Developed Dynamic Motion Transition (DMT), integrating data-driven and model-driven techniques.
• Employed data augmentation using physical laws and force-derived concepts.
• Utilized Conditional Temporal Transformer (CTT) for fine-level force change learning.
• Applied Dynamic Movement Primitives (DMP) for coarse-level motion creation.

Main Results:

• The proposed hybrid approach outperforms state-of-the-art methods.
• Achieved superior performance using a newly devised metric based on least action loss.
• Demonstrated effectiveness in generating realistic and lifelike human motion transitions.

Conclusions:

• The Dynamic Motion Transition (DMT) method offers significant advantages for virtual human animation.
• The hybrid approach enhances generative power and handles data limitations effectively.
• This work benefits various animation tasks including motion synthesis, control, tracking, and prediction.