电子测量与仪器学报

黎书玉,张旭东,孙锐,范之国.基于椭球拟合的点云旋转不变网络[J].电子测量与仪器学报,2022,36(9):111-117

基于椭球拟合的点云旋转不变网络

Point cloud rotation invariant network based on ellipsoid fitting

DOI：

英文关键词:point clouds deep learning classification segmentation rotation invariant lightweight

基金项目:国家自然科学基金（61876057，61971177）、安徽省重点研发计划科技强警专项(202004d07020012)、安徽省科技重大专项（202103a06020010）项目资助

作者	单位
黎书玉	1.合肥工业大学计算机与信息学院
张旭东	1.合肥工业大学计算机与信息学院
孙锐	1.合肥工业大学计算机与信息学院
范之国	1.合肥工业大学计算机与信息学院

Author	Institution
Li Shuyu	1.School of Computer and Information, Hefei University of Technology
Zhang Xudong	1.School of Computer and Information, Hefei University of Technology
Sun Rui	1.School of Computer and Information, Hefei University of Technology
Fan Zhiguo	1.School of Computer and Information, Hefei University of Technology

摘要点击次数: 644

全文下载次数: 1125

中文摘要:

点云携带丰富的几何信息,在计算机视觉领域具有独特优势。现有基于深度学习的三维模型分类与分割方法能有效识别固定视角下的物体,但在实际应用中,物体方向未知,使得点云描述存在旋转变换问题,极大影响网络的识别精度。针对点云的旋转性问题,提出一种轻量级的基于椭球拟合的旋转不变网络( point cloud rotation invariant network based on ellipsoid fitting, EFRI-N)。设计前置网络模块提取点云的旋转不变特征,包括椭球拟合和特征编码两个部分。通过椭球拟合算法标识原始点云的方向得到旋转不变坐标系,再将原始特征映射到该坐标系中,利用空间信息和角度信息进行编码得到点云的旋转不变特征;为了获取更丰富的几何信息,在分类分割网络中加入多层级的特征连接增强特征传播及复用,提高模型表征能力。采用国际知名公共数据集 ModelNet40 和 ShapeNet Parts 进行分类、分割实验,结果表明,该方法在处理旋转点云的任务中优于主流算法,网络识别精度提升了 1% ~ 62. 63%不等,并且网络的计算量和参数量都有着数量级的优势。满足单目标场景下对点云旋转不变性的使用要求,具有良好的应用价值。

英文摘要:

Point clouds have unique advantages due to its rich geometric information in computer vision field. Most of the existing point cloud classification and segmentation methods based on deep learning can identify the objects with canonical orientations. In real applications, there are problems of rotation transformation. In this paper, we propose a lightweight framework EFRI-N, namely, rotation invariant network of point cloud based on ellipsoid fitting, focusing on pointset rotation problems. We design a pre-network module to extract the rotation-invariant features. The ellipsoid fitting algorithm is used to identify the direction of the point clouds and obtain the rotation-invariant coordinate. Then the original features are mapped to the coordinate, and the rotation-invariant features were obtained by encoding the spatial and angular information. In order to obtain richer geometric information, multi-level feature connection is added to the network to enhance feature propagation and reuse. The classification and segmentation experiments are carried out by using the famous public datasets ModelNet40 and ShapeNet Parts. The results show that this method demonstrates better performance than state-ofthe-art methods in the task of processing rotating point cloud, and the network is improved by 1% ~ 62. 63%. Moreover, the computation amount and the number of parameters of the network have an order of magnitude advantage. It can meet the requirements of rotation invariance of point cloud in single object scenario and has good application value.

查看全文查看/发表评论下载PDF阅读器