Below are the 11:00 and 2:00 colloquium summaries for Friday, March 4th. Please see http://eecs.oregonstate.edu/graduate/colloquium/ for more coming up next week.
Friday March 4 11:00 - 11:50 AM Owen 106 Eliza Yingzi Du Assistant Research Professor Electrical Engineering Department U.S. Naval Academy One-Dimensional Iris Recognition and Performance Analysis of Partial Iris Recognition The iris is a valuable biometric for use in identification, with distinctive texture that remains stable throughout life. The striations, filaments, and rings that make up the iris pattern are very unique to each person. Because of its uniqueness to an individual, it can provide identification with very high confidence through large databases. Compared with other biometric features such as face and fingerprint, iris patterns are more stable and reliable. This talk will first explain what iris recognition is. Then we introduce the One-Dimensional Iris Recognition System. We will discuss the pros and cons of this system. Moreover, we will use this system for performance analysis of partial iris recognition. The One-Dimensional Iris Recognition System generates one-dimensional signatures to rank iris pattern similarities. It is translation, rotation, illumination and scale invariant. It is very different from traditional iris recognition systems which typically use two-dimensional iris patterns/codes/signatures for iris identification and recognition and require circular rotation for pattern matching purposes. Also, this approach uses the Du measure as a matching mechanism, and generates a set of the most probable matches instead of only the best match. Currently, iris recognition systems require a cooperative subject, who willingly stares into a camera for a few seconds. Under these conditions, the iris image obtained the maximum amount of iris information. On the other hand, for a noncooperative subject who may be facing away from the camera, only a portion of the iris information may be captured (a partial iris). Partial iris recognition algorithms would be very important in surveillance applications. Little research has been performed in this area. In this talk, we study the performance of the use of a partial iris for recognition. A partial iris performance analysis system based on the one-dimensional approach to iris identification is introduced. Using this system, we quantitatively analyze the relationship between the accuracy rate and the portion of the iris patterns. We will discuss some interesting experimental results. This talk is made for the general public (who have no idea about iris recognition) as well as for the professionals (in biometrics/iris recognition area). Biography Eliza Yingzi Du is an Assistant Research Professor and a founding member of the Center for Biometric Signal Processing at the United States Naval Academy (2003-present). She earned her Ph.D. in Electrical Engineering at the Univ. of Maryland, Baltimore County in 2003. Her research interest is in the field of digital signal/image processing and communication. She has invented a One-dimensional Iris Recognition System, created Du Measure (a hyperspetral similarity measurement method named after her), designed an automatic color thresholding scheme, developed an automatic text detection and restoration system, and designed the network protection and disaster prevention schemes for international submarine optical networks. Her research has been funded by the National Security Agency (NSA), Office of Research Lab (ONR), Naval Research Lab (NRL), and National Science Foundation (NSF). She has been serving as an active reviewer for over a dozen of journals. Friday March 4 2:00 - 2:50 PM Owen 102 Cristian Sminchisescu Fellow Artificial Intelligence Laboratory University of Toronto Learning and Inference Algorithms for Reconstructing 3D Human Motion in Monocular Video Many visual perception problems, including object tracking or scene reconstruction can be formulated as inference using non-linear models, defined over high-dimensional state spaces. Suboptimal modeling, model image matching ambiguities or occlusion, often lead to representations that are weakly constrained by the image data. Therefore, robust solutions typically involve estimating and propagating highly multimodal posterior state distributions. Trapping in unrepresentative peaks represents a significant difficulty for model state inference, and it is important to locate and, for dynamic scenes, track a set of representative configurations, over time. Learning models well-adapted for the task but with good generalization power is also a necessary step, in order to reduce the search complexity, and obtain reliable estimates. In the talk, I will present inference and learning algorithms, and demonstrate these on the problem of monocular tracking and 3D motion reconstruction of a 40-dimensional articulated human model. I will focus on both generative and discriminative algorithms, discuss their assumptions and propagation rules and show how these can be used independently, but also combined in robust trackers that can efficiently self-initialize and recover from failure. Biography Cristian Sminchisescu received an Engineering Diploma and a MSc. from the "Politehnica" University of Bucharest, in Computer Science and Systems Simulation and Modeling, in 1996 and 1997 respectively. He has obtained a MSc. in Computer Science from Rutgers University in 1999 and a PhD. in Computer Science and Applied Mathematics (with specialization in Imaging, Vision and Robotics), from INRIA (Institute National Politehnique de Grenoble) in 2002. He has been a fellow in the Artificial Intelligence Laboratory at the University of Toronto since January 2002, he holds a visiting faculty position at Rutgers and a research faculty position at TTI-C. He has been in the program committee for ECCV and CVPR 2004 and CVPR 2005 and serves as a regular reviewer for journals and conferences including SIGGRAPH, IEEE PAMI or ICCV. His research interests are in computer vision, machine learning and computer graphics. _______________________________________________ Colloquium mailing list [email protected] https://secure.engr.oregonstate.edu/mailman/listinfo/colloquium
