Research On Point Cloud Resampling Based On Deep Learning

As the difficulty of obtaining point cloud data continues to decrease,as well as the development of related post-processing algorithms in recent years,point clouds have been used more and more widely.Due to technical limitations,point clouds acquired by existing equipment often have some problems.For example,the distribution of the scanned point cloud is uneven,and there are a lot of redundant points and noisy points in the point cloud.In order to solve these problems,researchers have done a lot of work on point cloud resampling algorithms.Point cloud resampling refers to optimizing the position and distribution of a point cloud to improve its effect in subsequent applications.It can be divided mainly into two directions: upsampling and downsampling.With the rapid development of the deep learning technology,in recent years,there have been a series of attempts to apply deep learning to point cloud resampling tasks.At present,most of the resampling methods based on deep learning ignore the difference of saliency of each point in the point cloud,and cannot adjust the distribution of the point cloud as a whole.This paper has carried out research on adaptive resampling of point clouds,and proposed two algorithms based on deep learning,so that the resampled point clouds have different densities where the saliency is different.Specifically,the work of this article can be divided into the following two aspects:(1)We proposed a point cloud downsampling method based on point cloud saliency prediction and optimization.This method first uses a neural network to predict the saliency of the input point cloud,then selects the initial downsampling points based on the saliency of each point,and finally uses another network to optimize the initial point cloud position and provide a normal estimation.In order to train the neural network we designed,we also generated a dataset containing various label data such as point normals and saliencies.The experimental results on this dataset prove that our proposed method outperforms other existing downsampling methods in mesh distance metric and local uniformity metric.(2)We designed a point cloud resampling network based on a sampling matrix,which can generate adaptive sampling results of any number of points.The network first uses edge convolution and our proposed T-softmax structure to learn a sampling matrix,and then obtains the target distribution of the output point cloud through the sampling matrix.After that,the network extracts the local features from the input point cloud and the target distribution at the same time,and generates the final resampling result.We also proposed a new point cloud saliency algorithm,which uses both point cloud and mesh information for saliency label generation in our dataset.Subsequent ablation experiments show that our saliency algorithm performs better than other existing methods.Experimental results prove that our network is more effective than other current upsampling and downsampling methods in the task of resampling.