Testing face recognition systems

Abstract This paper discusses some of the difficulties which arise when reporting computer face recognition testing results. Examples are given, from the authors' own work, of factors that are not usually discussed which can have a significant bearing on the reported recognition rate. It is shown that a popular and apparently rigorous approach to testing, where the choice is between a fixed number of faces, each of which appears in each of a fixed number of different conditions, can produce unrepresentative results. The paper raises more general questions, and makes recommendations about the type of information that should be included in order to provide greater confidence in reported recognition rates.