An important problem in face recognition is the design of the feature space which represents the human face. Various feature sets have been and are continually being proposed for this purpose. However, there exists no feature set which gives a superior and consistent recognition performance on various face databases. Concatenating the popular features together and forming a high dimensional feature space introduces the curse of dimensionality problem. For this reason, dimensionality reduction techniques such as Principal Component Analysis is utilized on the feature space. In this study, first, some of the popular feature sets used in face recognition literature are evaluated over three popular face databases, namely ORL , UMIST , and Yale . Then, high dimensional feature space obtained by concatenating all the features is reduced to a lower dimensional space by using the Minimal Redundancy Maximal Relevance  feature selection method in order to design a generic and successful feature set. The results indicate that mRMR selects a small number of features which are satisfactory and consistent in terms of recognition performance, provided that the face database is statistically stable with sufficient amount of data.