Abstract
Generating 3D point clouds is challenging yet highly desired. This workpresents a novel autoregressive model, PointGrow, which can generate diverseand realistic point cloud samples from scratch or conditioned on semanticcontexts. This model operates recurrently, with each point sampled according toa conditional distribution given its previously-generated points, allowinginter-point correlations to be well-exploited and 3D shape generative processesto be better interpreted. Since point cloud object shapes are typically encodedby long-range dependencies, we augment our model with dedicated self-attentionmodules to capture such relations. Extensive evaluations show that PointGrowachieves satisfying performance on both unconditional and conditional pointcloud generation tasks, with respect to realism and diversity. Severalimportant applications, such as unsupervised feature learning and shapearithmetic operations, are also demonstrated.