A comparative study of automatic face verification algorithms on the BANCA database

The performance of different face identity verification methods on BANCA database is compared. As part of the comparison, we investigate the effect of representation on different approaches to face verification. Two conventional dimensionality reduction methods, namely the Principal Component Analysis and the Linear Discriminant Analysis are studied as well as the use of the raw image space. The results of the comparison show that when the training set size is limited, a better performance is achieved using Normalised Correlation method in the LDA space while Support Vector Machine classifier is superior when a large enough training set is available. Moreover, the SVM is almost insensitive to the choice of representation. However, a dimensionality reduction can be beneficial if constraints on the size of the template are imposed.