Analyzing the impact of police mugshots in face verification for crime investigations

During photo-signaling procedures, national police forces routinely collect two pictures (commonly known as mugshots), i.e., the frontal picture and the right profile picture, together with fingerprints and personal information of a subject, for various purposes, ranging from releasing documents to registering criminals. Thus, there is a clear connection between face recognition and this data collection procedure. Despite Automated Fingerprint Identification Systems (AFISs) are being extended by integrating face recognition techniques to provide image analysis capabilities, there is a lack of research in understanding to which extent such techniques are effective in identifying a known subject when only the two standard images of police databases are available as reference samples. This paper makes a step in such direction by comparing the performance of two different Convolutional Neural Networks (CNNs), pretrained on the VGGFace2 dataset, in a face verification task using the Surveillance Cameras Face (SCFace) database. Specifically, we organized subsets of mugshots taken from different angles in the range [-90°, 90°], where -90° is an image showing the left profile of a subject, 0° is the frontal picture, and 90° is the right profile. We used each subset in a face verification task with images taken from security cameras, trying to understand if the use of more than two mugshot pictures, taken from different points of view, has a positive impact in the identification of suspected subjects during crime investigations. This study might be a first step to assess if a revision of the mugshot collection procedure during photo-signaling is worth the necessary investments, in terms of picture acquisition and storage.