Deep Feature Extraction And Weakly-supervised Learning Of 3D Point Clouds

Recently,with the advance of 3D scanners such as LiDARs and RGB-D cameras,3D data has become increasingly available.At the same time,the numerous applications in different fields,such as autonomous driving,robotics,and reverse engineering,has attracted greater attention in understanding and analyzing 3D data.3D data can be represented in different formats,including depth images,meshes,3D voxels and point clouds.As a commonly used format,point clouds have become the preferred representation for many scene understanding applications,because it does not need any discretization and retains the original geometric information in 3D space.In recent years,deep learning technology has dominated many research areas,such as computer vision and natural language processing.However,due to the inherent disorder,irregularity and magnanimity of point clouds,deep learning on 3D point clouds is still facing some bottlenecks,especially the weakly-supervised learning of point clouds.Weakly-supervised learning involves two difficulties:feature extraction and label sparseness.This thesis mainly focused on the influence of point cloud disorder and irregularity on feature extraction,and the influence of magnanimity on label sparseness.And we conducted research in the following three aspects:1.In view of the disorder problem of point clouds,this thesis formulated a multilevel aggregation method of deep point cloud features based on dynamic graphs.Based on the spatial characteristics of 3D points,this method explored the correlation of different levels of point clouds sequentially,and coupled the relationship features in a nonlinear way,improving the effectiveness of feature extraction.2.To address the problem of irregular distribution of point clouds,this thesis introduced a point cloud feature extraction method based on bidirectional learning.Firstly,through two-direction iterations of feature extraction and point cloud shifting,combined with the multi-level aggregation feature enhancement,this method constructed a multilevel aggregation enhanced point cloud deep feature extraction network based on bidirectional learning.This provided a backbone network foundation for weaklysupervised learning of point clouds.3.In view of the problem of label sparseness,this thesis proposed a weaklysupervised learning framework of point clouds based on contrastive learning.Based on the point cloud feature extraction network of research result 2,this method introduced multi-dimensional contrastive learning,and made use of semantic equivalence of transformation,pseudo-label constraints and class feature orthogonality to improve weakly-supervised learning.Based on the public evaluation datasets,such as ModelNet40,ShapeNet Parts,S3DIS and ScanNet,we conducted performance comparison experiments,and proved the effectiveness and the advance of the weakly-supervised learning method proposed in this thesis.