Modeling Human Motion-Capture Data for Creativity

Abstract

Human motion-capture data can be represented, modeled, and generated through computational techniques. This thesis explores representations and strategies for querying, interpolating, and sequence modeling of motion-capture data. We employ spectral analysis of motion capture data to facilitate the query and comparison of movements, and identify target features for interpolation. We train a decoder-only transformer model on text-encoded motion-capture data, which we fine-tune for dance generation and movement classification. Our core contributions are defining interpolation and language model training procedures for generating motion-captured dance.