Arbitrary-scale Upsampling For Point Cloud Via Self-attention Mechanism

With the maturity of point cloud acquisition technology,more and more researchers are attracted to the field of point cloud.By studying traditional point cloud algorithms,it is found that most point cloud problems tend to use uniform and dense inputs.As a result,researchers have proposed point cloud sampling(point cloud super-resolution)to generate a more dense set of points from sparse,uneven,noisy inputs.Most importantly,this point set needs to fit the input geometric surface and be evenly distributed to fit the human eye’s visual perception habits.Traditional point cloud sampling methods are often optimized based on constraints such as local surface smoothing,which makes it difficult to adapt to complex model structures.With the rapid development of depth learning,people are trying to deal with the problem in this new method.For example,PU-Gan,MPU,Dis-Up,etc.handle the single-scale upsampling problem and achieve good results.However,the single-scale model at a time can only handle one task,which is inefficient and costly.Although the Mate-Up model solves the problem of sampling at arbitrary scale,it ignores global consistency and results in a hollow model.To address these issues,an end-to-end network is constructed at this paper.The main work of this paper is summarized as follows:(1)In this paper,a new method is constructed to deal with the problem of arbitrary-scale upsampling by means of joint training.In addition,this paper also designs a sampling rate coding module to encode different sampling rates and further guide the network to distinguish different sampling rates.(2)By studying the self-attention mechanism,this paper proposes a new joint residual attention module.The module encodes the spatial attention and channel attention of global and local information,and retains the important geometric feature points of the model.(3)This paper also provides a detailed data set processing scheme for arbitrary scale upsampling,and describes in detail the steps of data set selection,sampling,segmentation,multi-scale down sampling and so on.At the end of this paper,the proposed arbitrary scale upsampling method based on attention is applied to the model with sharp features and the smooth surface model respectively.It is found that the results are more uniform and the modeling effect is better than the current method.Finally,from the comparison of experimental quantitative indicators and visual effects,it is concluded that the method solves that the current method will lead to holes,and can generate arbitrary scale,locally uniform and globally consistent point cloud sampling results.