Abstract: The aim of this paper is to propose a strategy thatuses data fusion at three different levels to graduallyimprove the performance of an identity verificationsystem. In a first step temporal data fusion can beused to combine multiple instances of a single expertto reduce its measurement variance. If systemperformance after this first step is not good enough tosatisfy the end-user's needs, one can improve it byfusing in a second step result of multiple expertsworking on the same modality. For this approach towork, it is supposed that the respective classificationerrors of the different experts are de-correlated.Finally, if the verification system's performance afterthis second step is still not good enough, one will beforced to move onto the third step in which performancecan be improved by using multiple experts working ondifferent modalities. To be useful however, theseexperts have to be chosen in such a way that adding theextra modalities increases the separation in themulti-dimensional modality-space between thedistributions of the different populations that have tobe classified by the system. This kind of level-basedstrategy allow to gradually tune the performance of anidentity verification system to the end-user'srequirements while controlling the increase ofinvestment costs. In this paper results of severalfusion modules will be shown at each level. Allexperiments have been performed on the same multi-modaldatabase to be able to compare the gain in performanceeach time one goes up a level. !42
展开▼