3D point clouds are widely used in immersive telepresence, cultural heritage reconstruction, geophysical information systems, autonomous driving, and virtual/augmented reality. Despite rapid development in 3D sensing technology, acquiring 3D point cloud data with high spatial and temporal resolution and complex geometry/topology is still time-consuming, challenging, or costly. This talk will present our recent studies on computational methods (i.e., deep learning)-based 3D point cloud reconstruction, including sparse 3D point cloud upsampling, temporal interpolation of dynamic 3D point cloud sequences, and adversarial 3D point cloud generation.