PhD Oral Exam - Tianhao Xie, Computer Science

When studying for a doctoral degree (PhD), candidates submit a thesis that provides a critical review of the current state of knowledge of the thesis subject as well as the student’s own contributions to the subject. The distinguishing criterion of doctoral graduate research is a significant and original contribution to knowledge.

Once accepted, the candidate presents the thesis orally. This oral exam is open to the public.

Abstract

Machine learning has reshaped computer vision and graphics by enabling data-driven approaches to 3D synthesis and editing from text, images, and other forms of high-level supervision. Generative models, neural representations, and differentiable optimization have greatly expanded the flexibility of 3D content creation, allowing complex scenes and objects to be reconstructed, generated, and manipulated in ways that were difficult to achieve with traditional methods alone. However, despite their strong representational power, learning-based methods often remain limited by weak geometric control, ambiguous supervision, inefficient parameterization, and insufficient consistency in 3D space.

This thesis bridges classical graphics methods and modern machine-learning techniques for more controllable and efficient 3D synthesis and editing. Rather than treating graphics priors and learning-based models as competing paradigms, it explores how graphics methods such as regularization energies, deformation models, adaptive parameterization, and function reparameterization can be integrated into modern machine learning priors. These graphics-inspired components provide useful structural guidance that improves controllability, stability, efficiency, and geometric fidelity of the reconstructed and edited results. Depending on the task, they either regularize learning-based priors for more reliable 3D editing or enhance neural representations for more efficient and expressive 3D reconstruction.

Overall, this thesis argues that classical graphics methods remain valuable in modern learning-based frameworks. By revisiting and adapting classical graphics techniques within modern neural frameworks, we establish a novel path toward more reliable and effective 3D synthesis and editing.