摘要
点云和视图2种模态的全局特征融合已经被证明有效,为了进一步挖掘细粒度的局部特征关系和模态间的互补关系以提升模型性能,提出了一种新的网络架构。该架构包含2个核心模块:设计了局部特征融合模块(LFM),通过特征矩阵的转置变化,集成2种模态在不同层次的局部特征,而特征互补增强模块则利用元素级的简单运算,获取模态间的区别性信息,并以此为根据量化出权重系数,最后加权增强特征,形成更强大的形状描述符。在ModelNet10和ModelNet40数据集上的实验结果表明,该网络在效率和性能上实现了平衡,并且在三维形状识别方面取得了先进的成果。
The global feature fusion of point cloud and view modalities has been proven effective.To further explore fine-grained local feature relationships and complementary relationships between modalities and improve the performance of the model,we propose a new network architecture,which consists of two core modules.First,a Local Feature Fusion Module(LFM)is designed,integrating the local features of two modalities at different levels through transposed changes in the feature matrix.The feature complementarity enhancement module utilizes element level simple operations to obtain discriminative information between modalities and employs them as a basis to quantify weight coefficients.Finally,the feature is weighted and enhanced to form stronger shape descriptors.A large number of experiments conducted on the dataset ModelNet10/40 show the network reaches a balance between efficiency and performance and dilivers superior performances in 3D shape recognition.
作者
但远宏
王志浩
金毓
DAN Yuanhong;WANG Zhihao;JIN Yu(College of Computer Science and Engineering,Chongqing University of Technology,Chongqing 400054,China)
出处
《重庆理工大学学报(自然科学)》
北大核心
2025年第1期125-131,共7页
Journal of Chongqing University of Technology:Natural Science
基金
重庆市科委重点攻关计划项目(2021CCB03)
南京理工大学重点实验室基金赞助项目(2022-JCJQ-LB-061-07)。
关键词
多模态
三维形状理解
深度学习
局部特征
multimodal
understanding 3D shapes
deep learning
local feature
