Multimedia Information Analysis (MIA) Laboratory

Our research focuses on multimedia signal processing – to develop efficient, robust, and secure systems to analyze, process, and communicate multimedia information. It is a challenging area as most multimedia applications require handling data at a very high rate, demand a truly realistic user experience, but are often difficult to analyze and interpret.

Our core objective is to develop a research program that balances two goals: (1) to develop prototype systems for the end users with properly-defined measures of success, and (2) to investigate the scientific problems behind by proposing new theories to enhance our fundamental understanding of these problems. From protecting privacy in multimedia to developing multimedia assistive technologies for autism, all our projects aim at bringing radical changes in technology and solving significant societal problems.

Here are some of our current research projects:

Autism and Virtual Mirror

Clipboard01This project is an interdisciplinary, integrated research and education program to develop novel technologies in manipulating mirror images, aimed at studying and enabling behavioral modeling of children with autism spectrum disorder (ASD). Central to the research is a “virtual-mirror” device that combines a network of calibrated depth and visual sensors to render a viewpoint-dependent dynamic view of an arbitrary-shaped virtual mirror on a room-size see-through display. Through multimodal and spatially-diverse sensors, the proposed system provides high-fidelity, non-intrusive capturing of eye gaze, facial expression, body pose, body movement, and other human behavioral patterns. New multimedia processing algorithms will be developed for transferring 2D and 3D physical appearances, as well as behaviors from a source individual to a target individual with limited target training data to be rendered on regular displays and the virtual mirror.

Privacy in Distributed Multimedia Processing

Clipboard03This research breaks this “efficiency barrier” of the classical cryptographical approach by investigating a new computational framework to combine distributed multimedia processing and homomorphic encryption. Building on recent results of our group, this work develops efficient encrypted-domain processing through optimal computational procedures, parallel computation in cipher-text, small-field manipulation and encrypted data compression. In addition, by exploiting the interplay between privacy and the perceptual nature of multimedia, this work develops a provably-secure tradeoff scheme between privacy and complexity to significantly reduce the complexity and bandwidth requirements of encrypted-domain processing. We demonstrate the usability of this new framework through novel applications in biometric matching, object detection, speech analysis and video surveillance.