Date of Award

Fall 1-1-2025

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Computer Science

First Advisor

Rushmeier, Holly

Abstract

Humans, with complex structures and diverse attributes, pose significant challenges for digital modeling. One fundamental difficulty arises from the heterogeneous nature of humans, necessitating expressive yet efficient representations to faithfully capture fine details across different components. Moreover, even minor imperfections in digital humans can be highly perceptible, often leading to issues such as the uncanny valley. With the growing interest in the metaverse and digital human avatars, developing robust and high-fidelity representations for digital humans has become a pressing problem in computer vision and graphics, with significant implications for down-stream applications such as telepresence, virtual agents, and visual effects. In this dissertation, we propose a series of neural representations to model the intricate characteristics of human motion, hair, and head. Our first work focuses on motion modeling, introducing an implicit neural representation that learns the spatio-temporal space of kinematic motions. Unlike previous work that represents motion as discrete sequential samples, our method models it as a continuous function over time, enabling the formulation of a motion generative model conditioned on temporal coordinates and latent variables. Our second work focuses on hair modeling, presenting a learned parametric representation of 3D human hair. The core of our formulation is a novel strand representation in the frequency domain, allowing the decomposition of 3D hairstyles into low- to high-frequency structures. The parameterization of these decomposed hair structures leads to our parametric hair model, with disentangled parameters to control the global hair structure and local curl patterns, respectively. Our last work focuses on head modeling, proposing an unconditional generative model for 3D human heads with composable hair and face. Our representation leverages deformable hair geometry to capture the geometric variations across different hairstyles, along with a 3D generative model that explicitly models the separation and correlation between hair and face, yielding a head generative model that supports compositionality. Extensive experiments are conducted to validate the architecture design of our neural representations, with applications designed to demonstrate their potential in serving as generic priors to solve task-agnostic problems, further showcasing the flexibility and superiority of our works.

Download

Share

COinS