3D Reconstruction based on GAT from a Single Image

Previous methods which generate 3D shape usually use volume or a point cloud. These 3D shapes would lose details and are not smooth on the surface. This paper aims to introduce an end-to-end deep learning architecture which can generate 3D shape in triangular mesh from a single image. The architecture uses the graph attention network to continuously deform the initial ellipsoid based on the features extracted from a single image. The deformation process is constantly refined. The method can produce high precision mesh mode with richer details.