Learning Geometric Transformation for Point Cloud Completion

Point cloud completion aims to estimate the missing shape from a partial point cloud. Existing encoder-decoder based generative models usually reconstruct the complete point cloud from the learned distribution of the shape prior, which may lead to distortion of geometric details (such as sharp structures and structures without smooth surfaces) due to the information loss of the latent space embedding. To address this problem, we formulate point cloud completion as a geometric transformation problem and propose a simple yet effective geometric transformation network (GTNet). It exploits the repetitive geometric structures in common 3D objects to recover the complete shapes, which contains three sub-networks: geometric patch network, structure transformation network, and detail refinement network. Specifically, the geometric patch network iteratively discovers repetitive geometric structures that are related or similar to the missing parts. Then, the structure transformation network uses the discovered geometric structures to complete the corresponding missing parts by learning their spatial transformations such as symmetry, rotation, translation, and uniform scaling. Finally, the detail refinement network performs global optimization to eliminate unnatural structures. Extensive experiments demonstrate that the proposed method outperforms the state-of-the-art methods on the Shape-Net55-34, MVP, PCN, and KITTI datasets. Models and code will be available at https://github.com/ivislabhit/GTNet.