Using 3D pose alignment tools in forensic applications of Face Recognition

Face recognition (FR) systems have traditionally been designed to work with strictly controlled images and to be as fully automated as possible. This is driven by the civilian and commercial application requirements for high throughput, ease of use, and convenience. However, recently some developers are beginning to explore alternative approaches in order to address the unique requirements for forensic applications and improve the performance of the technology in this area. Methods include: incorporating operators as an integral part of the system's processes to enhance images; correcting for pose and alignment and manually marking up or encoding "local features" that can be used by the algorithm. This approach is analogous to that employed in many conventional fingerprint applications for searching forensic samples from crime scenes. This paper reports on tests performed using an FR system that employs User Interface (UI) tools that allow an operator to correct for pose (3D pose correction and alignment) to improve the search performance for forensic data. The system was used to search test datasets comprised of operationally representative CCTV and police custody images. The results showed that for CCTV-to-Mugshot and Mugshot-to-CCTV searches the use of manual 3D pose correction and alignment can provide significant improvements in performance.