Research On Multi-Encoder Shape Completion Algorithm Of Unstructured Point Cloud

Unstructured point clouds are a representative shape representation of real-world scenes in 3D vision and graphics.Point clouds with incomplete shape inevitably arise,due to the equipment to capture unstructured point clouds and the way of data processing,which has a huge impact on the accuracy of object recognition and detection,surface reconstruction,volume estimation and other tasks.Traditional methods such as symmetry hypothesis and database matching are used to complete the shape completion of point cloud,which are sensitive to the noise level of point cloud data and have poor accuracy and universality.In recent years,with the introduction of PointNet,people can use neural network to directly process point cloud data and extract point cloud features.On this basis,point cloud generation methods based on deep learning have been widely used and achieved good results,including methods for point cloud shape completion.In view of the existing point cloud shape completion methods through 2D images representation and 3D voxelization of point clouds,which require the structure,topology,and annotation of the prior basic shape,a sparse-to-dense multi-encoder neural network structure is proposed in this thesis,which completes and optimizes the defective point cloud in a sparse-to-dense manner of two-stages.The first stage,the network generates a sparse but complete point cloud by a bistratal PointNet and a three-layer fully connected network,and the second stage,the network yields a dense and high-fidelity point cloud by encoding and decoding the sparse result in the first stage using PointNet++.The method combines the distance loss and repulsion loss,and designs a joint loss function to optimize the learning ability of the network to generate more uniformly distributed output point cloud closer to the ground-truth counterparts.Through experiments on the public dataset ShapeNet,qualitative and quantitative tests and analysis show that the method proposed in this thesis is suitable for end-to-end uniform shape completion of unstructured point cloud and obtain relatively accurate local details of point clouds,it can be directly implemented on the incomplete or even containing noise and occlusion point cloud without any transformation,makes no specific assumptions about the incompletion distribution and geometry features in the input point cloud.and it has good robustness to noise or large-area occlusion,and it is superior to the existing learning-based point cloud shape completion methods in terms of real structure recovery and uniformity.