西南交通大学学报2024,Vol.59Issue(5):1184-1193,10.DOI:10.3969/j.issn.0258-2724.20220682
基于注意力机制的三维点云模型对应关系计算
Correspondence Calculation of Three-Dimensional Point Cloud Model Based on Attention Mechanism
摘要
Abstract
The existing deep learning methods have low precision and poor generalization ability in calculating dense correspondence between non-rigid point cloud models.To address these issues,a novel method for calculating unsupervised three-dimensional(3D)point cloud model correspondence based on a feature sequence attention mechanism was proposed.Firstly,the feature extraction module was used to extract the features of the input point cloud model pair.Secondly,the transformer module learned context information by capturing self-attention and cross-attention and generated a soft mapping matrix through the correspondence prediction module.Finally,the reconstruction module reconstructed the point cloud model based on the obtained soft mapping matrix and used the unsupervised loss function to complete training.The experimental results on FAUST,SHREC'19,and SMAL datasets show that the average correspondence errors of this algorithm are 5.1,5.8,and 5.4,respectively,which are lower than those of the classical algorithms including 3D-CODED,Elementary Structures,and CorrNet3D.The correspondence between non-rigid 3D point cloud models calculated by the proposed algorithm has higher accuracy and stronger generalization ability.关键词
计算机视觉/对应关系/无监督/点云重构/注意力机制Key words
computer vision/correspondence/unsupervised/point cloud reconstruction/attention mechanism分类
信息技术与安全科学引用本文复制引用
杨军,高志明,李金泰,张琛..基于注意力机制的三维点云模型对应关系计算[J].西南交通大学学报,2024,59(5):1184-1193,10.基金项目
国家自然科学基金项目(42261067,61862039) (42261067,61862039)
兰州市人才创新创业项目(2020-RC-22) 兰州交通大学天佑创新团队项目(TY202002)的资助. (2020-RC-22)