LARGE-SCALE NEURAL 3D SCENE RECONSTRUCTION, RENDERING, AND BEYOND

This thesis advances neural 3D reconstruction into a scalable and robust process for real-world applications. While techniques like NeRF and 3D Gaussian Splatting (3DGS) show promise, they face challenges in scalability, robustness to imperfect camera poses, and consistency across distributed captures. To address these, we introduce a series of innovations. For scalability, AdaSfM enables distributed camera pose estimation by adapting to scene complexity. Building on this, DOGS, a distributed training framework for 3DGS, reduces computation and memory by an order of magnitude for large-scale scenes. For robustness, DBARF eliminates the need for accurate precomputed poses by jointly optimizing them with the NeRF, succeeding where traditional methods fail. Finally, DReg-NeRF ensures consistency by automatically aligning neural fields from different captures. Together, these contributions form a comprehensive framework that demonstrates: - Scalability: Distributed pose estimation and rendering for city-scale scenes. - Robustness: Pose-free rendering in challenging environments. - Consistency: Automatic alignment for unified reconstruction. Rigorous validation on diverse datasets confirms significant improvements in accuracy and efficiency, enabling the practical deployment of neural 3D reconstruction where traditional approaches fall short.

Author dc:creator

• CHEN YU

• Large Scale
• Neural Rendering
• Gaussian Splatting
• Neural Radiance Fields
• Structure from Motion
• 3D Reconstruction

Licence

• https://scholarbank.nus.edu.sg/bitstreams/5d8dabff-40be-46ee-8f17-2b8ffc3f4693/download