Animatable 3D Human Avatar Modeling From Multi-View Images
Guardado en:
| Publicado en: | PQDT - Global (2025) |
|---|---|
| Autor principal: | |
| Publicado: |
ProQuest Dissertations & Theses
|
| Materias: | |
| Acceso en línea: | Citation/Abstract Full Text - PDF Full text outside of ProQuest |
| Etiquetas: |
Sin Etiquetas, Sea el primero en etiquetar este registro!
|
| Resumen: | Animatable 3D human avatar modeling is a long-standing research problem in computer vision and computer graphics. The task aims to capture and synthesize realistic human appearances and movements from visual observations, thus enabling immersive and interactive experiences in virtual reality, augmented reality, telepresence, film production, and gaming. Despite significant advancements, current approaches still face critical challenges, such as limited capability in modeling dynamic textures, inability to perform real-time rendering of avatars, and difficulty in generalizing high-quality 3D human reconstruction and animation. This thesis addresses these challenges by proposing a series of novel techniques, paving the way for more accessible, efficient, and realistic digital human content creation. Specifically, our contributions are threefold:In the first part of the thesis, we present 4K4D++, a novel dynamic human reconstruction method leveraging a temporally continuous 3D Gaussian representation. By explicitly modeling temporal coherence and utilizing a lightweight image-based rendering appearance module, 4K4D++ significantly enhances reconstruction efficiency and captures fine-grained dynamic details, surpassing state-of-the-art methods in rendering quality and speed.In the second part, we introduce NBAvatar, a novel approach for human avatar reconstruction that enables both high-fidelity appearance rendering and real-time animation. Our key innovation lies in neural blend features, a pose-dependent feature representation, substantially improving texture fidelity and dramatically reducing computational overhead. Extensive evaluations on public datasets demonstrate that NBAvatar achieves comparable visual fidelity to state-of-the-art methods while providing significant speedups in both training and animation.Finally, we propose HumanRAM, a generalizable feed-forward framework for human reconstruction and animation from single-view or sparse-view inputs. HumanRAM integrates parametric human priors into the large reconstruction model, enabling high-quality human reconstruction and high-fidelity human animation. Comprehensive experiments validate that our approach significantly outperforms existing methods in novel view and novel pose synthesis, demonstrating robust generalization capabilities on real-world data. |
|---|---|
| ISBN: | 9798263314316 |
| Fuente: | ProQuest Dissertations & Theses Global |