Research Context

This project studies text-driven Human-Human Interaction (HHI) motion generation, where a model must generate two-person motion from a natural-language description.

• HHI motion generation is harder than single-person motion generation because the output must preserve contact, distance, timing, role, and event structure between two people.
• Existing HHI models often rely on a sentence-level text embedding (CLIP text embedding), which can weaken fine-grained instructions such as body part, direction, role, or action-specific details.
• I approached this problem through two related directions: word-level text-motion alignment and interaction-structure analysis.

My Contribution

I worked on fine-grained conditioning and interaction-representation analysis for HHI motion generation as a first-author KSC 2025 paper and graduation-thesis project.

• Proposed Word-Motion Cross Attention (WMCA), where flattened motion tokens attend directly to CLIP word-token embeddings instead of relying only on a single sentence-level text condition.
• Defined motion-derived interaction features describing how two people interact over time, including contact frequency, distance, relative-position change, movement synchronization, and motion-pattern similarity.
• Built an audit process to test whether these features could support HHI generation: predicting them from text, injecting them as generation conditions, ranking candidate motions, and checking where they did or did not explain interaction quality.

Research Artifacts

WMCA architecture

Quantitative evaluation

Feature audit — fixed-generator candidate selection

Output

This work produced a KSC 2025 first-author paper and a graduation thesis on fine-grained condition utilization in text-driven HHI motion generation.

• KSC 2025: Word-Motion Cross Attention for Fine-Grained Text-Guided Human-Human Interaction Motion Generation.