Marios Savvides is a Research Professor in the Electrical and Computer Engineering Department of Carnegie Mellon University and also in Carnegie Mellon's CyLab. Dr. Savvides is also the Director of the CyLab Biometrics Center. His research is in developing Biometric Identification technologies and algorithms that work under co-operative scenarios (, i.e. recognizing a person based on their face, iris, fingerprint,and palmprints) and also un-cooperative scenarios (using surveillance data to recognize a person). Savvides collaborates and works in joint projects with Prof. B.V.K. Vijaya Kumar and Prof. Pradeep Khosla.
His research in Biometrics has been mostly focused on Face Recognition and Iris Recognition, developing new technology that can achieve distortion tolerant face & iris recognition. The appearance of face images can vary due to a number of factors such as pose, expression and illumination. Thus Savvides has been researching in developing techniques such as advanced correlation filters that have built-in tolerance to such variations. In the iris field, intra-class variations include local deformations, focus blur and off-angle iris views.
Recently (the past year) he has been spearheading and leading our CMU efforts in the Face Recognition Grand Challenge (FRGC) and the Iris Challenge Evaluation(ICE) which are parts of NIST's efforts in evaluating and identifying key performance technologies in Face recognition and Iris Recognition. This is a project that he works jointly with Prof. B.V.K. Vijaya Kumar, infact they are the only two faculty in CMU participating in FRGC and FRVT (the Face Recognition Vendor Test 2006) and more remarkably, they are also participating in ICE too (that makes them the only group doing both in academia and industry!).
In the latest FRGC-Phase II, Savvides has been involved and leading the development of our novel recognition approach for tackling Exp 4 in FRGC-II data, yielding 72% verification @ 0.1% False Acceptance Rate(FAR) in the pure 1-1 FRGC matching protocol (these is based using all pairwise comparisons of target/query images on the FRGC dataset). This is a dramatic improvement in performance of the baseline PCA algorithm which yeilds only 12% verification @ 0.1 % FAR. Their algorithm ranked 2nd with the top being at 76% @ 0.1% FAR. However, in a practical scenario, they have more than one mug-shot of a suspect that they are looking for, when they use such information and train using all images of a person in the target set, they can boost performance in FRGC Exp4 to 92% @ 0.1% FAR using discriminant learning with support vector machines and their advanced correlation feature extraction method.
PhD, 2005, Electrical and Computer Engineering, Carnegie Mellon University
MS, 2000, Robotics, Carnegie Mellon University
BEng, 1997, Microelectronics Systems Engineering, University of Manchester Institute of Science and Technology