Abstract: In order to overcome the quality problem of reconstructed meshes caused by the restrictive factors such as sparse and uneven density or wrong normal vector of original point cloud data,  a three-dimensional reconstruction method of point cloud based on Countermeasure network is proposed by using the characteristics of weight sharing in adversarial neural network and adversarial training process.Firstly, the predictor is used to predict the offset of the edge of the mesh model, so as to obtain the displacement of each vertex, and perform the vertex redirection of the topology preservation to obtain a new mesh. Secondly, the point cloud classifier in the discriminator is used to extract the high-dimensional features of the original point cloud data and the sampled point set of the mesh surface, and the spatial perception discrimination based on the high-dimensional features is performed to distinguish the original point cloud from the samples point set data. Finally, the output data of the predictor and the discriminator are correlated by the adversarial training method, and the network model is optimized through multiple iterations to obtain a three-dimensional mesh model that meets the spatial characteristics of the point cloud. Experiments are carried out on different point cloud datasets, and the results are displayed by using MeshLab software. The results show that the method can reconstruct the 3D mesh model that meets the spatial information of point cloud, and overcome the mesh quality problems caused by poor point cloud data.

Key words: 3D reconstruction, deep learning, mesh deformation, adversarial network