PhD Defense: Motion Modeling for Humans and Dynamic Objects

Motion is fundamental to how humans and dynamic objects interact with the world. Modeling motion is therefore essential for understanding dynamic phenomena, from reconstructing motion in real-world observations to generating plausible new motions. However, motion does not occur in isolation: it depends on an entity's physical properties, how it is observed, and how it interacts with others. Capturing these different forms of context remains a key challenge in motion modeling.
In this dissertation, we study motion modeling across humans and general dynamic objects. First, we introduce a text-driven human motion synthesis framework that jointly predicts body shape and motion, enabling shape-aware motion generation across different body morphologies. Second, we develop a template-free framework for reconstructing arbitrary dynamic objects from casual monocular videos, combining generative priors, object-centric camera estimation, and long-term point tracking. Finally, we extend motion modeling to multi-person interactions by constructing a large-scale dataset from sitcom videos and developing a text-conditioned motion generation framework that models spatial relationships and temporal structure to generate coordinated multi-person motions. Together, these works address complementary challenges in motion modeling, spanning body shape, dynamic 3D reconstruction, and human interactions.