EEMCS EPrints Service
van Dam, C. (2017) From Image Sequence to Frontal Image: Reconstruction of the Unknown Face, A Forensic Case. PhD thesis, Univ. of Twente. CTIT Ph.D. Thesis Series No. 17-429 ISBN 978-90-365-4324-8
Full text available as:
Official URL: http://dx.doi.org/10.3990/1.9789036543248
Faces, we see them every day. People recognize faces from a distance without any problem. Have you ever wondered why we are doing such a great job at recognizing faces? Or do we fool ourselves by thinking that we can recognize faces well? We consider it an easy task to recognize our relatives and friends, even when we only view them in low resolution images. But when we see our relatives and friends at an unexpected location, it may take a while before we recognize them. And how many times do we even fail to recognize them? So, the surroundings and context are definitely important for people in recognizing other people’s faces. Often when e think we recognize someone, we rather recognize their clothing or haircuts, instead of their face. Our brains seem to build some model of a person, especially a person’s face, based on all encounters with this person15. A model that can be updated, but which can also give a bias towards the recognition in certain surroundings or situations. One way to test our abilities to recognize faces is to perform an experiment with unfamiliar faces. How well would we recognize people that we have only seen in an image or a video? For many people it is a difficult task to recognize an unfamiliar face independently from its surroundings. And how to deal with the possible bias towards people we have encountered that we have built up over the years? In order to avoid such a bias, we would search for an automated approach to compare faces. As a result of years of research and development automated face recognition systems perform very well on frontal facial images. State-of-the-art face recognition systems make fewer than thirty errors on every thousand images78. However, the recognition performance of face recognition systems degrades for facial images under pose56. And how to combine the result of multiple recognitions for image sequences? Especially uncontrolled situations with faces under pose are difficult to handle for automated face comparison systems. To handle multiple images at once, 3D based reconstruction methods for pose and illumination compensation are needed. This 1 compensation could be acquired by creating a reconstruction of the face using 3D face models, but how much bias towards the 3D face models is introduced in such a process? In forensic casework it is even more important to avoid any bias towards specific face models that are not part of the case data. Is there a way to avoid introducing bias towards any data that is not part of the facial data during a reconstruction procedure? The case of the reconstruction of the unknown face is a difficult one.
Export this item as:
To correct this item please ask your editor
Repository Staff Only: edit this item