Chen, Zhiqin2023-11-282023-11-282023-06-16https://diglib.eg.org:443/handle/10.2312/3543919PhD thesis.Deep learning has revolutionized the field of 3D shape reconstruction, unlocking new possibilities and achieving superior performance compared to traditional methods. However, despite being the dominant 3D shape representation in real-world applications, polygon meshes have been severely underutilized as a representation for output shapes in neural 3D reconstruction methods. One key reason is that triangle tessellations are irregular, which poses challenges for generating them using neural networks. Therefore, it is imperative to develop algorithms that leverage the power of deep learning while generating output shapes in polygon mesh formats for seamless integration into real-world applications. In this thesis, we propose several data-driven approaches to reconstruct explicit meshes from diverse types of input data, aiming to address this challenge. Drawing inspiration from classical data structures and algorithms in computer graphics, we develop representations to effectively represent meshes within neural networks. First, we introduce BSP-Net. Inspired by a classical data structure Binary Space Partitioning (BSP), we represent a 3D shape as a union of convex primitives, where each convex primitive is obtained by intersecting half-spaces. This 3-layer BSP-tree representation allows a shape to be stored in a 3-layer multilayer perceptron (MLP) as a neural implicit, while an exact polygon mesh can be extracted from the MLP weights by parsing the underlying BSP-tree. BSP-Net is the first deep neural network that is able to produce compact and watertight polygon meshes natively, and the generated meshes are capable of representing sharp geometric features. We demonstrate its effectiveness in the task of single-view 3D reconstruction. Next, we introduce a series of works that reconstruct explicit meshes by storing meshes in regular grid structures. We present Neural Marching Cubes (NMC), a data-driven algorithm for reconstructing meshes from discretized implicit fields. NMC is built upon Marching Cubes (MC), but it learns the vertex positions and local mesh topologies from example training meshes, thereby avoiding topological errors and achieving better reconstruction of geometric features, especially sharp features such as edges and corners, compared to MC and its variants. In our subsequent work, Neural Dual Contouring (NDC), we replace the MC meshing algorithm with Dual Contouring (DC) with slight modifications, so that our algorithm can reconstruct meshes from both signed inputs, such as signed distance fields or binary voxels, and unsigned inputs, such as unsigned distance fields or point clouds, with high accuracy and fast inference speed in a unified framework. Furthermore, inspired by the volume rendering algorithm in Neural Radiance Fields (NeRF), we introduce differentiable rendering to NDC to arrive at MobileNeRF, a NeRF-based method for reconstructing objects and scenes as triangle meshes with view-dependent textures from multi-view images. MobileNeRF is the first NeRF-based method that is able to run on mobile phones and AR/VR platforms thanks to the explicit mesh representation, demonstrating its efficiency and compatibility on common devices.en3D shape, mesh, and representation3D reconstruction from voxels, point clouds, and imagesmachine learningThesis