An Auto-encoding model for 3D object surface reconstruction

As the development of deep learning continues to mature, automatically generating object surface shapes from 3D point cloud data has gradually become a research hotspot in the field of 3D reconstruction. Recent methods rely on the encoder-decoder structure to create a point cloud feature representation from the input data, and then learn the mapping of the 2D parameter space to the 3D surface to decode the features. AtlasNet network is representative of this method, which can generate high-resolution surfaces. Based on the principle of parametric 3D surface reconstruction of Auto-encoding model, this paper proposes a new Auto-encoding network for 3D surface reconstruction. First, PointNet++ as the new encoder is used to extract the local features of the point cloud. After the local features are grouped into larger units, higher-dimensional features are obtained through the network, and combined with the local features of each point, the global feature of point cloud is obtained after pooling. The sampling points of the 2D square are then embedded on the global features as input to the decoder, and finally a surface with high resolution can be generated.