In order to promote evaluation of pattern recognition algorithms using publicly available datasets and standard performance assessment methodology, a competition for the best face verification algorithms will take place in conjunction with the ICPR 2004 conference.
The authors of the best performing algorithm of the competition will receive a financial prize that will be presented at the Conference Dinner. The performance of all submitted algorithms will be summarised in a presentation given at the conference.
The tests will be carried out using the English images acquired from the BANCA database. A brief description of the database is presented at AVBPA2003 (Fourth International Conference on Audio and Video-based Biometric Person Authentication) is available on-line from here. Data from the database are available at the cost of distribution from here. (in order to promote the competition the cost of the database has been halved until December 2003 for academic institutions)
The experiments will be carried according to the BANCA protocol using the Pooled(P) configuration. Files specifying which images should be used for training and testing can be found in this directory. There are three separate parts to the competition.
Images will be supplied when you receive the BANCA data set, which have already been localised (by hand) and geometrically. The resulting resolution of the images is 55x51 pixels and 8-bit grey-scale and contains just face pixels. Information about the geometric noralisation can be found here. These images should then be used to train and test your authentication system. To enter this part of the contest the false acceptance and false rejection rates have to be submitted to us on configuration P of the BANCA protocol.
Full video resolution colour images, each containing a face will be supplied. All participants must use an automatic method of localisation for the at least the test phase of the protocol. You may use manual localisation in the training phase. Eye co-ordinates for the training images will be supplied. Again, the error rates (i.e. FA and FR) are required for entrance into this part of the competition.
In this part of the test, the algorithms will be tested on sequestered data. Participants must submit a fully trained, fully automatic verification system in a form of an executable running on one of the specified architectures (PC Linux, Sun Solaris, PC WindowsXX, etc.) taking two inputs: an image file and an client identification number and return 0 or 1 indication of a client (genuine claim) or 0 an impostor (false) identity claim.