Continuity is widely regarded as a fundamental characteristic of the material world. This thesis focuses on the concept of continuity within various 3D visual learning problems. The primary objective is to explore how 3D visual learning algorithms can effectively leverage the advantages of (quasi-)continuum. To address this overarching research question, our investigation is divided into two main directions. Firstly, we delve into continuity in 3D data representations. Recent advancements in implicit neural representations have yielded impressive outcomes by encoding 3D signals using continuous neural functions. We examine learning-based implicit neural representations from a generalisation perspective, specifically focusing on how latent-coded 3D implicit functions generalise across a range of shape geometries. Our analysis involves tracking local surface point trajectories alongside the global latent interpolation, enabling us to gain insights into the hierarchical functionality in implicit neural layers. Furthermore, we extend implicit neural representations to accommodate the symmetry of geometric transformations. Secondly, we explore the utilization of continuity in learning algorithms for 3D vision tasks. We investigate the role of continuity in data distribution, label space, and architecture design for point cloud classification and RGB-D segmentation tasks. Our research demonstrates the benefits of maintaining continuity in these learning algorithms. By examining these two aspects, our aim is to shed light on the potential of continuity to enhance the performance and capabilities of 3D visual learning. This thesis contributes to a deeper understanding of how continuity can be effectively leveraged in 3D visual learning, ultimately paving the way for future advancements.