MOST: Motion Diffusion Model for Rare Text via Temporal Clip Banzhaf Interaction.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Generating human motion from text prompts is challenging, especially with rare language.
• Existing methods often suffer from coarse-grained matching and overlook semantic details due to motion redundancy.

Purpose of the Study:

• To introduce a novel MOtion diffuSion model via Temporal clip Banzhaf interaction (MOST).
• To address the limitations of previous approaches in text-to-motion generation, particularly for rare prompts.

Main Methods:

• MOST utilizes a retrieval stage with temporal clip Banzhaf interaction to quantify textual-motion coherence at the clip level.
• This enables fine-grained text-to-motion clip matching, reducing redundancy.
• A motion prompt module in the generation stage uses retrieved clips for semantically consistent motion.

Main Results:

• MOST achieves state-of-the-art performance in both text-to-motion retrieval and generation.
• Evaluations demonstrate significant improvements, especially for rare and complex language prompts.
• Quantitative and qualitative results confirm the model's effectiveness in addressing previous challenges.

Conclusions:

• MOST effectively generates human motion from language prompts, even rare ones.
• The novel temporal clip Banzhaf interaction significantly improves text-motion coherence and reduces redundancy.
• This approach advances the field of AI-driven human motion synthesis.