Evaluating Deep Neural Networks for Face Recognition with Different Subsets of Mugshots From the Photo-Signaling Procedure

Whereas Face Recognition with Convolutional Neural Networks (CNNs) is considered a mature technology to the point that, in addition to applications such as smartphone unlocking and passport verification, it is used in already existing decision support systems for crime investigations, there still are open challenges in Pose-Invariant Face Recognition (PIFR). Specifically, there is a lack of research in understanding how subsets of mugshots different from the frontal and right profile pictures routinely collected by police forces during the photo-signaling procedure might impact on the Face Recognition accuracy in security camera videos recorded 'in the wild'. To this end, we compare two well-known CNNs for Face Recognition, namely VGG16 and ResNet50, on the Face Recognition from Mugshots DataBase (FRMDB), specifically designed to evaluate the performance in Face Recognition using different subsets of mugshots. With respect to our previous research, we collect more general results, testing with more identities on additional security camera videos.