Lagrangian Motion Fields for Long-Term Motion Generation.

Area of Science:

• Computer Graphics
• Artificial Intelligence
• Animation

Background:

• Long-term motion generation is difficult due to limitations in current framewise representations.
• Existing methods overlook temporal dynamics, leading to redundancy and hindering realistic long-term motion synthesis.

Purpose of the Study:

• Introduce Lagrangian Motion Fields for effective long-term motion generation.
• Overcome the limitations of framewise representations by integrating spatial and temporal motion dynamics.

Main Methods:

• Represent joints as Lagrangian particles with uniform velocity over short intervals.
• Condense motion representations into "supermotions" analogous to superpixels.
• Develop a lightweight, versatile solution that bypasses neural network preprocessing.

Main Results:

• Achieve enhanced efficiency, superior generation quality, and greater diversity in motion synthesis.
• Demonstrate effectiveness in long-term music-to-dance and text-to-motion generation.
• Showcase adaptability for infinite motion looping and controlled motion generation.

Conclusions:

• Lagrangian Motion Fields offer a novel and effective paradigm for long-term motion generation.
• The approach seamlessly integrates spatial and temporal information, outperforming existing methods.
• Its versatility and efficiency make it suitable for a wide range of animation and motion synthesis applications.