Progressive Point Cloud Generating by Shape Decomposing and Upsampling

How to generate realistic and precise 3D point clouds of different shapes has become a hot topic for many researchers. However, due to the complexity and discreteness of point cloud data, generating high-resolution complete point clouds from implicit vectors is still very challenging. It is difficult to obtain high-resolution points in one step. Besides, it is also hard to recover local details of a shape. To tackle these two challenges, we propose a new point cloud generation network. It learns to decompose the partial structures and extract the local information from them. We use this local information to guide the generation of local shapes. To control the quality of point clouds during the procedure, we progressively generate dense point clouds through iterative upsampling, in which the next step is performed based on the priority of the previous step. To be specific, We first generate a coarse point cloud and then decompose it into multiple local parts. Each of them is then mapped to a latent local feature. For each local area, we gradually improve the resolution of the shape by multiple upsampling layers. Each upsampling layer combines local features, global features, and point-wise features to predict two offsets for each point to increase the number of points. Our network can progressively generate uniformly multi-resolutions point clouds, with realistic and detailed local structures. Experiment results demonstrate that our model’s generation ability is better than other GAN-based generation methods.