Surface reconstruction from unstructured 3D point clouds is a fundamental challenge, especially in the presence of sparse, noisy, or incomplete data. To address the limitations of traditional meshing and signed distance function methods, we propose im...
Surface reconstruction from unstructured 3D point clouds is a fundamental challenge, especially in the presence of sparse, noisy, or incomplete data. To address the limitations of traditional meshing and signed distance function methods, we propose implicit neural frameworks: Neural Shortest Path (NSP) and Neural Eikonal Solver (NES). NSP learns both the distance function and its gradient via a magnitude-direction decomposition, ensuring convergence in the Sobolev space $H^1$ and enabling robust surface localization. NES approximates solutions to the eikonal equation using two formulations: a viscosity-regularized model (Vis-NES) and a Cole-Hopf-based linearization (CH-NES), both implemented as implicit neural representations. We also introduce surface extraction algorithms tailored to each framework, enhancing mesh quality without relying on traditional methods such as marching cubes. Extensive experiments on synthetic and real-world datasets demonstrate the accuracy, robustness, and theoretical grounding of the proposed methods.