Evaluating the Complexity of Databases for Person Identification and Verification

For the development and evaluation of methods for person identification, verification, and other tasks, databases play an important role. Despite this fact, there exists no measure whether a given database is sufficient to train and/or to test a given algorithm. This paper proposes a method to ``grade'' the complexity of a database, respectively to validate whether a database is appropriate for the simulation of a given application. Experiments support the argumentation that the complexity of a data set is not equivalent to its size. The ``first nearest neighbor'' method applied to image vectors is shown to perform reasonably well for person identification, respectively the mean square distance for person verification. This suggests to use them as a minimal performance measure for other algorithms.

Related material