Research On Super-resolution Point Cloud Generation Method Based On Deep Learning

With the popularization of depth sensors,the research around 3D point clouds has received great attention from academics and the industry.In some popular studies such as autonomous driving,virtual reality,simultaneous localization and mapping,point cloud plays an indispensable role,and many data-driven methods are introduced into the field of point cloud research,which has achieved excellent results in the above research tasks.Super-resolution point cloud generation refers to the input of sparse point clouds through the algorithm to obtain a dense high-quality point cloud,which is essentially similar to the image super-resolution problem.However,there is little work published for the super-resolution point cloud generation task,due to the unordered form of representation and the inherent ambiguity of ground truth.In view of the above research status,this paper explores the task of super-resolution point cloud generation based on deep learning.The main work and contributions can be summarized as follows:1.In this paper,a two-stage super-resolution point cloud generation framework based on local interpolation and neural network re-adjustment is proposed.In the first stage of the framework,this paper proposes an interpolation method based on local features to upsample the input point cloud,and obtain the initial upsampling result while ensuring the consistency of point cloud density.In the coarse to fine stage,this paper designs a neural network based on the outer product to further re-adjust the interpolated point cloud coordinates to complete the conversion of the point cloud from coarse to fine.The experimental results obtained on the datasets used in the existing methods prove that the proposed framework has smaller error and more accurate point cloud results than the existing methods.The Poisson reconstruction results of the point cloud also demonstrate the effectiveness of the proposed method.2.In order to generate super-resolution point cloud in a more efficient way,this paper proposes a patch-based end-to-end super-resolution point cloud reconstruction method for the shortcomings of the previous proposed framework.In the previous method,the point cloud coordinates are adjusted point by point and the geometric features on different scales are not considered.In the improved network,the multi-level feature extraction is used to capture the density information and geometric information of different regions.After the local features and global features of the patch are gathered,the reconstructed high-resolution point cloud is finally obtained.In addition,the performance of the end-to-end network transformed into a point cloud auto-encoder is further explored.The experimental results verify the effectiveness of the end-to-end point cloud reconstruction method.