Multi-view 3D Reconstruction with Self-attention

Different CNN-based algorithms have been used to reconstruct different 3D model data. Even so, there are still many defects in the reconstruction method for a single 3D model. The purpose of this paper is to learn the potential features of the 2D image, use the possible features to predict the complete 3D point cloud model, compare the output entire point cloud with the actual 3D point cloud and modify the parameters of the trained network model to achieve better reconstruction effect. In addition, this paper proposes a self-attention method that is widely used in 2D images and applies it to 3D reconstruction. After feature extraction of 2D images, the attention-weighted sum will be used to extract features of each group of images for subsequent 3D point prediction. The proposed method achieved an overall average Chamfer-L2 Distance for 13 categories of 7.03 for 24 2D views, which compared favorably to PSGN and 3D-R2N2.