Please use this identifier to cite or link to this item:
|Scopus||Web of Science®||Altmetric|
|Title:||A scalable algorithm for learning a Mahalanobis distance metric|
|Citation:||Proceedings of 9th Asian Conference on Computer Vision (ACCV'09), 2009 / H. Zha, Rin-ichiro Taniguchi and S. Maybank (eds.) PART III; pp.299-310|
|Publisher:||Springer Berlin Heidelberg|
|Publisher Place:||New York|
|Conference Name:||Asian Conference on Computer Vision (9th : 2009 : Xi'an, China)|
|Junae Kim, Chunhua Shen and Lei Wang|
|Abstract:||A distance metric that can accurately reflect the intrinsic characteristics of data is critical for visual recognition tasks. An effective solution to defining such a metric is to learn it from a set of training samples. In this work, we propose a fast and scalable algorithm to learn a Mahalanobis distance. By employing the principle of margin maximization to secure better generalization performances, this algorithm formulates the metric learning as a convex optimization problem with a positive semidefinite (psd) matrix variable. Based on an important theorem that a psd matrix with trace of one can always be represented as a convex combination of multiple rank-one matrices, our algorithm employs a differentiable loss function and solves the above convex optimization with gradient descent methods. This algorithm not only naturally maintains the psd requirement of the matrix variable that is essential for metric learning, but also significantly cuts down computational overhead, making it much more efficient with the increasing dimensions of feature vectors. Experimental study on benchmark data sets indicates that, compared with the existing metric learning algorithms, our algorithm can achieve higher classification accuracy with much less computational load.|
|Rights:||© Copyright 2009 IEEE – All Rights Reserved|
|Appears in Collections:||Computer Science publications|
Files in This Item:
There are no files associated with this item.
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.