Research Of Data Augmentation In Point Cloud Deep Learning

In recent years,the research of 3D deep learning has become a current research hotspot.It has been widely used in fields such as autonomous driving and robotics.The number of samples of the 3D model data set in the current research is limited,and the scale cannot be compared with the 2D data set.This is also a problem that plagues most researchers.The lack of data sets in deep learning tasks will directly lead to bottlenecks in the performance of the model,and data augmentation methods are proposed in this process.However,in the field of 3D point cloud deep learning,there is not much research on data augmentation technology.This article mainly studies the data augmentation algorithm for 3D point cloud data.The proposed method can fundamentally solve the problem of insufficient data.In previous research work,there have been many mature algorithms for 2D image data augmentation.Combining the mature ideas of 2D data augmentation algorithms with the characteristics of 3D point cloud data,data augmentation algorithms suitable for 3D data have been designed.The main contributions of this paper are as follows:(1)Based on the occlusion-based data augmentation algorithm,this paper analyzes the shortcomings of the three-dimensional random cropping algorithm.According to the two-dimensional data augmentation algorithm GridMask,a three-dimensional cropping method 3DGridMask that can balance over cropping and insufficient cropping is designed.(2)Based on the hybrid data augmentation algorithm,this paper analyzes the principle of Mixup and solves the defect that the Mixup algorithm cannot be directly applied in three-dimensional space,and combines the theory of Augmix to design a single-sample mixing method 3DAugmix,which can avoid direct mixing The problem of generating invalid sample points.(3)Based on the data augmentation algorithm of the class activation map,this paper designs a class activation map generation algorithm suitable for 3D point cloud data.Combining this method to design the 3DSnapmix algorithm,the label of the augmentation sample can be accurately obtained.This paper uses PointNet++ and RS-CNN models on the ModelNet dataset to conduct experiments to prove the effectiveness of the three-dimensional data augmentation algorithm proposed in this paper,and analyze the interpretability and execution efficiency of these several data augmentation algorithms.