Generating Distance-Aware Human-to-Human Interaction Motions From Text Guidance

Area of Science:

• Computer Vision
• Artificial Intelligence
• Animation

Background:

• Text-driven 3D human motion synthesis is advancing, but generating realistic multi-person interactions remains difficult.
• Current methods often neglect crucial physical constraints like explicit spatial distances between interacting body parts.

Purpose of the Study:

• To develop a novel system for generating realistic, text-controlled multi-person 3D animations.
• To address the limitations of existing models in capturing plausible inter-character spatial dynamics.

Main Methods:

• Proposes InterDist, a masked generative Transformer model operating in a discrete state space.
• Decomposes two-person motion into independent single-person motions and a separate interaction distance sequence.
• Utilizes a VQ-VAE for encoding and a bidirectional masked generative Transformer for modeling, conditioned on text, with a cross-modal interaction module.

Main Results:

• The InterDist model successfully generates motions that are semantically aligned with text prompts.
• Ensures the preservation of plausible inter-character spatial distances during animation.
• Sets a new benchmark for text-driven multi-person interaction generation.

Conclusions:

• InterDist offers a significant advancement in text-driven multi-person motion generation by explicitly modeling spatial interactions.
• The approach enhances both the semantic accuracy and physical plausibility of generated animations.
• This method paves the way for more realistic character interactions in virtual environments.