• FVC2006 is a contest accompanying the International Conference on Biometrics 2006 and it is organized by Prof. Josef Kittler from University of Surrey, who has been organizing face evaluations for many years in Europe.

• Unlike the FERET and FRVT, FVC2006 focuses on face verification rather than face identification. It is conducted on a database of 296 subjects, with each subject 12 images of varying lighting, expression, pose, and time intervals.

• The results of FVC2006 are announced on the ICB2006, and a summary paper on the contest is published in the FVC2006 proceeding.

• There are totally 9 algorithms have participate in the contest mainly from the IDIAP and University of Surrey.

• The following figure summarizes the comparison between our results and the best results from other participants (as well as the best results from the previous AVBPA03 contest on the same dataset using the same evaluation protocol.)

• For more information, please refer to the ICB06 paper.

Our method, Gabor Feature based Multiple Classifier Combination (CAS-GFMCC) (BMVC05), is an ensemble learning classifier based on the manipulation of Gabor features with multiple scales and orientations. The basic procedure of CAS-GFMCC is described as follows: First, face images are aligned geometrically and normalized photometrically by using region-based histogram equalization (AMFG2003). Then, Gabor filters with 5 scales and 8 orientations are convoluted with the normalized image and the magnitude of the transform results are kept for further process. These high dimensional Gabor features, with a dimension of 40 times of the original normalized face images, are then adaptively divided into multiple groups (BMVC05, FG04-1, AMFG05). And then, for each feature group, one classifier is learnt through Fisher discriminant analysis, which will result in an ensemble of classifier. These classifiers are combined together through fusion strategy.

  For automatic evaluation case, AdaBoost-based methods are exploited for both the localization of the face (FG04-2) and facial landmarks, i.e., the two eyes (IWBRS05). And to enhance the robustness of the system to mis-alignment caused by the imprecise eye positions, eye perturbation has been used to augment the training set (FG04-3). Please refer to the linked papers for more details of our methods.

   In addition, face image re-lighting techniques are exploited to adapt the method for more robustness to the face images with complex illumination (named by CAS-GFMCC-L) (ISCAS2002).

You are also suggested to refer to other related paper published on CVPR05, ICCV2005, IJPRAI2005. For the list of all our publications, please click here.