Diverse Motion In-Betweening From Sparse Keyframes With Dual Posture Stitching.

Area of Science:

• Computer Graphics
• Artificial Intelligence
• Animation

Background:

• Existing in-betweening methods often require multiple input frames, which are not always available.
• Current techniques can produce results lacking diversity, failing to meet artistic needs.

Purpose of the Study:

• To develop a method for generating diverse, high-quality character transitions using only two input frames (start and target).
• To address the limitations of existing in-betweening techniques regarding input requirements and result diversity.

Main Methods:

• Proposes a bi-directional motion generation and stitching scheme.
• Utilizes two adversarial autoregressive networks (Conditional Variational Autoencoders) to generate forward and backward transitions.
• Introduces a novel stitching loss for midway fusion of generated motions.

Main Results:

• Achieves higher motion quality compared to existing methods.
• Generates more diverse transition results.
• Demonstrates effectiveness on LaFAN1, Human3.6m, and AMASS datasets.

Conclusions:

• The proposed method successfully generates high-quality and diverse character transitions from minimal input.
• The bi-directional approach and flexible midway stitching offer significant advantages over prior work.