Local CCA alignment and its applications

详细信息	查看全文 | 推荐本文 |

• 作者：Lai Wei^a ; ^{weilai@shmtu.edu.cn} ; ^{weily105@hotmail.com}  ; [Author Vitae] ; Feifei Xu^b ; ^{xufeifei1983@hotmail.com}  ; [Author Vitae]
• 关键词：Canonical correlation analysis (CCA) ; Dimensionality reduction ; Data visualization ; Pose estimation
• 刊名：Neurocomputing
• 出版年：2012
• 期刊代码：95_09252312
• 类别：cp
• 出版时间：15 July, 2012
• 卷：89
• 期：Complete
• 页码：78-88
• 文件大小：814 K

摘要

Canonical correlation analysis (CCA) is a well-known technique for finding the correlations between two sets of multi-dimensional variables. It projects both sets of variables into a lower-dimensional space in which they are maximally correlated. One popular use of CCA is for dimensionality reduction. CCA can be regarded as a linear subspace approach for one view of an object set (e.g. ) which is directed by another view of the object set (e.g. ). However, if the correlations between and are nonlinear, CCA may fail to reveal the latent structures of . In this paper, we propose a new nonlinear dimensionality reduction algorithm, called local canonical correlation analysis alignment (LCCA). In LCCA, CCA is implemented on patches of an object set to get the local low-dimensional coordinates of ( is a patch of ), then the local coordinates are aligned to obtain the global low-dimensional embeddings of . Furthermore, in order to solve out-of-sample problems, a linear version of LCCA (LLCCA) algorithm is also developed. Different from LCCA, LLCCA is not only suitable for training samples but also for testing samples. Experiments for data visualization and pose estimation show that LCCA and LLCCA are superior to the related algorithms.