Research On Monocular 3D Reconstruction Technology Based On Voxel And Point Cloud

For human beings living in a 3D world,various 3D shapes can be seen everywhere.Compared with a 2D image that is reduced by one dimension,the 3D shape contains more information.Due to low cost,flexibility and other reasons,most of the camera systems in real life are monocular systems,so there are abundant monocular 2D image resources on the Internet.The goal of 3D reconstruction based on monocular vision is to restore the missing dimensions of the corresponding 3D shape from a single 2D image,and infer the 3D geometric structure of objects and scenes.The 3D reconstruction based on monocular vision is a basic task in computer vision,and it has a wide range of applications in robotics,computer-aided design,virtual reality and augmented reality.Due to the illposedness of 3D reconstruction,the overall reconstruction quality of the voxel-based 3D reconstruction method is not high,and the reconstruction quality of multiple categories is poor.The 3D reconstruction method based on point cloud is more effective than the method based on voxel.However,the point cloud is a kind of unstructured data,and the convolution structure often used in voxel reconstruction is difficult to fully utilize in point cloud reconstruction.Aiming at the problem that the overall reconstruction quality of the current voxelbased 3D reconstruction method is not high,a new 3D voxel reconstruction network based on shape layer is designed according to the good conversion relationship between shape layer and voxels with resolution of 32 × 32 × 32,converting the prediction problem of 3D voxel into the prediction problem of 2D depth maps.Experiments on the reconstruction network on the Shpae Net dataset have achieved better results compared with other voxelbased 3D reconstruction methods.Although the results of the 3D voxel reconstruction network based on the shape layer are improved compared with the previous methods,it still cannot solve the problem that many categories do not perform well on the evaluation indicators.This is a common defect of the voxel-based methods.Because the voxel-based 3D reconstruction method has the problem of poor reconstruction effects on multiple categories,a point cloud representation is selected to replace the voxels,which changes the technical route.However,the point cloud-based method also has the problem that it is difficult to use convolution.Aiming at these two problems,a new 3D point cloud reconstruction network based on deformed network is designed.In this dissertation the specific point feature,style feature and global feature of the point cloud are researched,a new point cloud feature extraction method is proposed.A fully connected point cloud deformation network and a Graph X-based multi-resolution point cloud deformation network are designed.Experiments were performed on the point cloud-based 3D reconstruction network on the Shpae Net data set.When the chamfer distance is used as the evaluation index,the point cloud reconstruction network obtained the best results.When Io U is used as the evaluation index,although the optimal results are not obtained,the reconstruction result Io U evaluation is less than 60% of the categories is successfully reduced from 6 to 1.Partially solves the problem that the voxel-based method has many types of poor reconstruction quality evaluation.