To further prove the validity of our approach, Table 3 presents cross-database results in which one database is used for training and the other for testing. The cross-database results show that the local-DNN models can generalize well to a different database at the expense of a small performance penalty. It should be emphasized that our cross-database results are better than the only previously published cross-database results presented in [4] where a 81.02% and 89.77% are obtained when testing with the Gallagher’s and LFW respectively.