Princeton University

School of Engineering & Applied Science

Short codes in large-scale identification systems: theoretical and practical considerations

Sviatoslav Voloshynovskiy, University of Geneva
Engineering Quadrangle B205
Thursday, March 13, 2014 - 12:30pm to 1:30pm

<strong>Abstract:</strong> In this talk, I present a statistical framework for the analysis of the performance of Bag-of-Words (BOW) identification systems and link it with coding and information-theoretic perspectives. The BOW framework has been widely used in content search systems, biometric applications and more recently in multimedia security applications. Modern BOW based systems can easily handle large-scale search or recognition problems, even on mobile phones. The BOW approach is based on the construction of a visual alphabet or dictionary based on the clustering of low-level features such as discriminative and robust descriptors. These low-level features represent a sort of short codes extracted from the content. Often binary, random and in the order of hundred bits, these short codes do not individually achieve satisfactory performance. However, under the proper joint decoding many experimental systems demonstrate quite impressive results. Nowadays, the design of existing BOW-based systems is based on memory/complexity considerations in view of the large-scale nature of the search problem. It includes a lot of heuristics and engineering, where performance is mostly evaluated by testing on (large) databases, and empirically compared for different descriptor classes and encoding algorithms. The presentation aims at establishing a better understanding of the impact of different elements of BOW systems such as the robustness of descriptors, accuracy of assignment and encoding, descriptor compression and pooling and finally decoding. We will demonstrate some information-theoretic limits of BOW system performance. We also study the impact of geometrical information on the BOW system performance and compare the results with different pooling strategies. The proposed framework can also be of interest for a security and privacy analysis of BOW systems. The experimental results on real images and descriptors confirm our theoretical findings.
<strong>Biography:</strong> Sviatoslav Voloshynovskiy (IEEE Senior Member'11) received the Radio Engineer degree from Lviv Polytechnic Institute, Lviv, Ukraine, in 1993 and the Ph.D. degree in electrical engineering from the State University “Lvivska Polytechnika,” Lviv, Ukraine, in 1996. From 1998 to 1999, he was a visiting scholar with the University of Illinois at Urbana-Champaign. Since 1999, he has been with the University of Geneva, Switzerland, where he is currently an Associate Professor with the Department of Computer Science and head of the Stochastic Information Processing group. His current research interests are in the information-theoretic aspects of digital data hiding, content fingerprinting, physical object security, stochastic image modeling and machine learning. He has co-authored over 200 journal and conference papers in these areas and holds twelve patents. He is Associate Editor for IEEE Transactions on Information Forensics and Security, Eurasip Journal on Information Security and an elected member of the IEEE Information Forensics and Security Technical Committee (2011-2013) where he is area chair in information-theoretic security. He has served as a consultant to private industry in the above areas.