Approaches for High-Fidelity Novel View Synthesis: From Casual Captures to Novel Views

This thesis investigates how the selection, quality, and geometric arrangement of views influence the fidelity of novel view synthesis. The first part focuses on reconstructive approaches based on Neural Radiance Fields and 3D Gaussian Splatting, introducing methods for image deblurring, pose independent frame selection, and view utility estimation. The second part studies generative approaches based on multiview diffusion models, with a particular focus on camera geometry, conditioning view arrangement, sparse view novel view synthesis, and 360° panorama generation. The results demonstrate that both input quality and camera geometry play a central role in reconstruction and generation fidelity, while also showing that video trained multiview diffusion models can be transferred zero shot to new view synthesis tasks without task specific fine tuning.

Examination Committee

• Prof. Vincent Lepetit, École des Ponts ParisTech (Reviewer)
• Prof. Mounir Kaaniche, Université Sorbonne Paris Nord (Reviewer)
• Dr. Frédéric Dufaux, CNRS Research Director, CentraleSupélec, Université Paris-Saclay (Examiner)
• Prof. Sylvie Le Hégarat, Université Paris-Saclay (Examiner)