Abstract: For the task of point cloud completion, it is always difficult to maintain the original shape structures effectively due to its missing of inherent features of the input point cloud. To this end, a structure-preserving point completion network is proposed that can incorporate the input incomplete point cloud. The proposed network adopts an encoder-decoder structure, and the encoder uses a multi-layer perception and a maximum pooling layer to obtain its feature codeword of the input 3D shape. The generated feature codeword can thus be performed a folding operation using four 2D grids to obtain a rough repairing result, which can be fused with the input point cloud. The final completed point cloud can be obtained by the iterative farthest point sampling (IFPS) operation. Compared with the completion results on ModelNet40 dataset, the average distance error of our proposed network is 11%-53% lower than that of the existing completion networks, whilst the average error of our network will be 15%-28% lower than that of the existing methods on ShapeNet dataset, and 59%-70% lower than that of the existing methods for point cloud shapes with fine structures. Experimental results demonstrate that our completion network can maintain the structural features of the input shapes effectively while repairing the missing parts of the underlying point clouds and is robust to different degrees of data loss. Compared with the existing networks, our completion network always has smaller distance error and uniform distribution of sampling points.