Learning Support Vectors for Face Authentication: Sensitivity to Mis-Registrations (0)
| Venue: | In Proceeding of ACCV'00. British Machine Vision Conference 9 |
| Citations: | 2 - 2 self |
BibTeX
@INPROCEEDINGS{Jonsson_learningsupport,
author = {K. Jonsson and J. Kittler and J. Matas},
title = {Learning Support Vectors for Face Authentication: Sensitivity to Mis-Registrations},
booktitle = {In Proceeding of ACCV'00. British Machine Vision Conference 9},
year = {}
}
Bookmark
 |
 |
 |
 |
 |
 |
OpenURL
Abstract
The paper studies Support Vector Machines (SVMs) in the context of face authentication. We present an evaluation of the impact of registration errors on the SVM performance. The results obtained by randomly perturbing the groundtruth eye coordinates are compared to those of a fully automatic registration technique based on a robust form of correlation. In the latter method, mis-registrations occur when the optimisation of the similarity function fails and the search terminates in a local maximum. The results of the comparison show that, unless prior knowledge about the problem-specific invariances is incorporated into the SVM training algorithm, the performance is likely to suffer. Nevertheless, we present competitive results obtained on a large face database of 295 subjects and we show that the combined system based on the robust correlation for face registration and the SVMs for verification compares favourably with two other methods based on the dynamic link architecture. Keywords:...
