采用交替优化策略的三维模型兴趣点提取算法  

Detecting Points of Interest on 3D Model via Alternating Optimization Strategy

作　　者：舒振宇 易顺[3] 杨思鹏 刘予琪 隆威 金海容 辛士庆 吴双卿 Shu Zhenyu;Yi Shun;Yang Sipeng;Liu Yuqi;Long Wei;Jin Hairong;Xin Shiqing;Wu Shuangqin(School of Computer and Data Engineering,NingboTech University,Ningbo 315100;Ningbo Institute,Zhejiang University,Ningbo 315100;School of Mechanical Engineering,Zhejiang University,Hangzhou 310027;State Key Laboratory of CAD&CG,Zhejiang University,Hangzhou 310058;College of Information Science&Electronic Engineering,Zhejiang University,Hangzhou 310027;Polytechnic Institute,Zhejiang University,Hangzhou 310015;School of Computer Science and Technology,Shandong University,Qingdao 266237;School of Information Science and Technology,NingboTech University,Ningbo 315100)

机构地区：[1]浙大宁波理工学院计算机与数据工程学院,宁波315100 [2]浙江大学宁波研究院,宁波315100 [3]浙江大学机械工程学院,杭州310027 [4]浙江大学CAD&CG国家重点实验室,杭州310058 [5]浙江大学信息与电子工程学院,杭州310027 [6]浙江大学工程师学院,杭州310015 [7]山东大学计算机科学与技术学院,青岛266237 [8]浙大宁波理工学院信息科学与工程学院,宁波315100

出　　处：《计算机辅助设计与图形学学报》2022年第7期1095-1107,共13页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(61872321,62172356);浙江省自然科学基金(LY22F020026,LQ17F030002);宁波市自然科学基金(2017A610108);宁波市“科技创新2025”重大专项(2020Z005,2020Z007,2021Z012).

摘　　要：针对三维模型的兴趣点提取问题,提出一种基于交替优化的全监督检测算法.第1步,利用多种特征描述符对人工标注好的三维模型进行特征提取,得到每个顶点的特征向量,将其作为神经网络的输入;第2步,使用双调和距离场为模型表面顶点赋予概率标签,并将顶点标签值作为神经网络的输出;第3步,通过神经网络学习输入特征与输出标签之间的复杂映射关系;第4步,将训练后的神经网络对训练集进行预测,并把兴趣点提取结果与人工标签进行对比,根据对比差异进一步优化顶点标签值,然后将顶点标签值作为输出、顶点特征向量作为输入,继续优化神经网络.将第3步和第4步重复多次进行交替优化,最终得到一个较优的神经网络模型.在公开数据集SHREC 2011上的实验结果表明,由于采用了交替优化的策略,所提算法在三维模型表面兴趣点提取的关键评价指标FNE和FPE上均优于传统算法,准确率实现了平均11个百分点以上的提升.A supervised 3D points of interest(POI)detection algorithm is proposed based on alternating optimi-zation.Firstly,the geometric features of a 3D shape are calculated from several hand-crafted feature descriptors and used as the input of the neural network.Secondly,the biharmonic distance field is utilized to assign a label to each vertex,which is regarded as the neural network’s output.Thirdly,the complex mapping relationships be-tween the feature vectors and the labels are learned through the neural network.Fourthly,predictions are made on the training set using the trained neural network.The differences between the predicted points of interest and ground truth are compared to further optimize the vertices’labels,which are then used as the output to train the neural network.The third and fourth steps for alternating optimization are repeated for several times,and a neural network is finally obtained.The experimental results on the SHREC 2011 dataset show that,due to the alternate optimization strategy,our algorithm is better than the traditional methods in the key evaluation indicators FNE and FPE,and the accuracy of proposed algorithm has achieved an average improvement of more than 11%.

关 键 词：三维模型 兴趣点 交替优化 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]