Learning-based methods for 3D reconstruction of indoor scenes

3D scene reconstruction is an foundamental and important topic in 3D computer vision, with many applications such as mixed/augmented reality, autonomous navigation, and robotics. This thesis is dedicated to the analysis and development of learning-based algorithms aimed at advancing the performance of 3D indoor reconstruction. Particularly, it tackles two key challenges across various reconstruction approaches, i.e., the use of data priors and the design of scene parameterization, striving for improved performance and broader applicability. In detail, Chapters 2, 3, and 6 explore the integration of various data priors for reconstruction: imaging prior and boundary priors are integrated in Chapter 2, geometric priors are exploited in Chapter 3, and multi-view consistency and 3D local smoothness priors are utilized in Chapter 6. In contrast, Chapters 4 and 5 focus on scene parameterization, where Chapter 4 introduces a novel scene representation and Chapter 5 targets accurate density parameterization in neural implicit representations.