自监督非等距三维模型簇对应关系计算方法
Self-Supervised Non-Isometric 3D Shape Collection Correspondence Calculation Method
吴衍 1杨军 2张思洋3
作者信息
- 1. 兰州交通大学电子与信息工程学院 兰州 730070;福建技术师范学院大数据与人工智能学院 福清 350300
- 2. 兰州交通大学电子与信息工程学院 兰州 730070;兰州交通大学测绘与地理信息学院 兰州 730070
- 3. 兰州交通大学电子与信息工程学院 兰州 730070
- 折叠
摘要
针对现有非等距三维模型簇对应关系计算方法准确率较低且泛化能力较差的问题,文中提出采用深度内外特征对齐算法的自监督非等距三维模型簇对应关系计算方法.首先,利用DiffusionNet直接学习原始的三维模型特征,获取具有鉴别力的特征描述符.然后,使用深度内外特征对齐算法计算非等距模型对之间的对应关系,采用局部流形调和基作为模型本征信息,并结合笛卡尔坐标等外部信息,实现内外部信息对齐一致,以无监督的方式自动生成对应结果.最后,构建非等距模型簇的加权无向图,根据相似几何模型存在内在相关性的原则,设计自监督多模型匹配算法,不断增强模型图最短路径的循环一致性,获得最优的非等距三维模型簇对应关系.实验表明,文中方法对应关系测地误差较小,结果较准确,能处理模型自身对称性影响对应关系计算的问题,具有良好的泛化能力.
Abstract
Aiming at the problem of low accuracy and poor generalization ability in existing non-isometric 3D shape collection correspondence calculation methods,a self-supervised non-isometric 3D shape collection correspondence calculation method using deep intrinsic-extrinsic feature alignment algorithm is proposed.Firstly,discriminative feature descriptors are obtained by directly learning the original 3D shape features through DiffusionNet.Then,the deep intrinsic-extrinsic feature alignment algorithm is employed to compute correspondences between non-isometric shapes.Consistency between internal and external information is realized by utilizing local manifold harmonic bases as intrinsic information of the shapes and integrating external information such as Cartesian coordinates.Consequently,correspondence results are generated automatically in an unsupervised manner.Finally,a weighted undirected graph of non-isometric shape collections is constructed.Based on the principle of inherent correlation among similar geometric shapes,a self-supervised multi-shape matching algorithm is designed to continuously enhance the cycle-consistency of the shortest path in the shape graph,and thus optimal correspondences for non-isometric 3D shape collections are obtained.Experimental results demonstrate that the proposed method achieves small geodesic errors in correspondences with accurate results,and effectively deals with the symmetric ambiguity problem with good generalization ability.
关键词
对应关系/非等距模型簇/自监督/深度学习/内外特征对齐Key words
Correspondence/Non-Isometric Shape Collection/Self-Supervised/Deep Learning/In-trinsic-Extrinsic Feature Alignment引用本文复制引用
基金项目
国家自然科学基金(42261067)
福建省自然科学基金(2022J01972)
中央引导地方科技发展资金项目(2021)(2021-51)
国家自然科学基金(61773301)
出版年
2024
NETL
NSTL
万方数据