Abstract: Point cloud completion estimates the complete shape given incomplete point cloud, which is a crucial task as the raw point cloud measurements suffer from missing data. Most of previous methods for point cloud completion share the encoder-decoder structure, where the encoder projects the raw point cloud into low-dimensional latent space and the decoder decodes the condensed latent information back into the list of points. While the low-dimensional projection extracts semantic features to guide the global completion of the missing data, the unique local geometric details observed from partial data are often lost. In this paper, we propose a shape completion framework that maintains both of the global context and the local characteristics. Our network is composed of two complementary prediction branches. One of the branches fills the unseen parts with the global context learned from the database model, which can be replaced by any of the conventional shape completion network. The other branch, which we refer as a Symmetry-Aware Upsampling Module (SAUM), conservatively maintains the geometric details given the observed partial data, clearly utilizing the symmetry for the shape completion. Experimental results show that the combination of the two prediction branches enables more plausible shape completion for point clouds than the state-of-the-art approaches.