| With the development of sensor acquisition technology,the medium of computer perceiving the world has gradually expanded from 2D data,e.g.images,to 3D,e.g.depth images and 3D point cloud.Among them,3D point cloud is collected from the surface of an object in 3D space and is usually described as a set of discrete points distributed in 3D space,describing the shape and geometric of the objects.The 3D point cloud in the distributed space is an unstructured data representation.Generally,the point cloud data has the characteristics of displacement invariance and rotation invariance,while the description of 3D scenes with point cloud data has the characteristics of scale invariance,occlusion free and category imbalance.Recently,the deep learningbased methods has become one of the state-of-the-art approaches for solving problems in the fields of computer vision and graphics due to its strong feature learning ability.Therefore,applying the deep learning-based method to the classification and segmentation tasks of 3D point cloud and carrying out advanced intelligent perceptual processing of point cloud data is the basis for systematic analysis and understanding of 3D environment using computers,which is also the focus and hot topic of academic research and industrial application.This dissertation focuses on the research of intelligent perception tasks,e.g.,classification and segmentation for 3D point cloud with the help of deep learning and technologies,around the intrinsic properties of real-world 3D point cloud objects and scenes,Starting from analyzing and summarizing the advantages and disadvantages of existing research methods,in order to solve the key points and difficulties of theoretical innovation and practical application,the dissertation proposes several new depth learning frameworks and model training methods.Firstly,the dissertation draws on the graph attention mechanism of graph neural network to extract stable consistent global feature representation for rotated point cloud data,and realizes accurate classification and segmentation of point cloud data with arbitrary rotation angle.Subsequently,by explicitly considering the semantic segmentation of large-scale point cloud scenes,an efficient attention module that can handle massive point clouds is proposed,and the idea of mutual information is used to achieve more accurate segmentation within semantic segmentation of scenes.Finally,focusing on the difficulty of point by point annotation of real point cloud scenes,a 3D point cloud weakly-supervised semantic segmentation framework based on a novel contrastive regularization is proposed.Meanwhile,focusing on the long-tail distribution in real scenes,a decoupling optimization training framework is further proposed for class imbalanced semi-supervised semantic segmentation tasks of point cloud scenes.To verify the effectiveness and efficiency of these methods,extensive experiments are conducted on several publicly available 3D point cloud datasets with significant improvement on performance,while maintaining high computational efficiency and low training cost.In summary,the main contributions and innovations of this article can be elaborated from the following aspects:(1)Research on rotation invariance of point cloud based on graph convolution.To allow the model to focus on the rotation invariance of the point cloud data,a graph-based mechanism is applied to deal with the invariant geometric features of point cloud data.A new graph convolution operator and a hierarchy aware graph pooling operator are proposed to achieve arbitrary rotations.Meanwhile,local geometric features and global consistent features of point cloud data are extracted,and classification and segmentation models are proposed for point cloud data with arbitrary rotations.It has achieved superior performance gains of 2.9%,2.4% and0.4% for shape classification,part segmentation and semantic segmentation of 3D point cloud data with arbitrary rotation angles.(2)Research on point cloud scenes segmentation based on the attention mechanism and mutual information.To compensate for the problem of accurate recognition of objects in large-scale scenes containing massive point clouds,spatial attention and channel attention modules are proposed.At the same time,considering class-imbalanced problem,as well as intra-class inconsistency and interclass indistinction of the semantic segmentation task,two new loss functions are proposed by using the idea of mutual information and voxel occupancy.The method achieves an average performance improvement of 2.5% on indoor largescale 3D point cloud scene dataset.(3)Research on weakly-supervised point cloud segmentation based on hybrid contrastive regularization.To address the huge labeling cost in large-scale point cloud semantic segmentation,a novel hybrid contrastive regularization framework for point cloud weakly-supervised semantic segmentation is proposed.Further,a dynamic point cloud augmentor to generate diversity and robust sample views is designed,whose transformation parameter is jointly optimized during training.Through extensive experiments,the framework achieves significant performances against the state-of-the-art methods on several large-scale point cloud scene datasets with limited annotation,which achieves an average performance gain of 2.4 % and1.0 %,respectively.(4)Research on class-imbalanced semi-supervised point cloud segmentation based on decoupling optimization.To overcome the problem of class-imbalanced problem in semi-supervised learning for point cloud scenes segmentation.A new decoupling optimization framework are proposed,which disentangles feature representation learning of backbone and classifier in an alternative optimization manner to shift the bias decision boundary effectively.In particular,two-round pseudo label generation strategy is employed and a multi-class imbalanced focus loss is introduced.Extensive experiments demonstrate the effectiveness of the framework,which gain of 6.8% on the indoor point cloud scene dataset S3 DIS with limited annotations. |
PDF Full Text Request