Praveen Seshadri, Assistant Professor, Computer Science Department

There is increasing need to be able to query, analyze and summarize large volumes of heterogeneous data. Two factors are critical to the success of query processing services for heterogeneous information: (1) functionality – the ability to model and manipulate complex data types from heterogeneous repositories, and (2) quality of service – the ability to predict the quality and timeliness of the results is in many ways more important than the absolute performance. Our research is developing the basic database technology needed to provide these abilities.

This project is building a next-generation object-relational database server, called PREDATOR, that has three goals: (1) to provide extensible support for complex data types like images, video, audio, and time-series; (2) to manipulate such complex data in combination with traditional database queries more efficiently than current technology permits; and (3) to extend these techniques to access heterogeneous data repositories from within the query processing system. PREDATOR is based on a novel database architectural design we have developed – an enhancement of abstract data types called E-ADTs. The initial results from the support of time-series and image data show query processing improvements of orders of magnitude over current technologies. We are currently developing support for several complex data types including audio, video, images, documents and mathematical structures.

Ramin Zabih, Assistant Professor, Computer Science Department

The Web's biggest advantage is the enormous amount of information available on it, but the Web's biggest drawback is the difficulty of finding the desired information. Search engines provide a partial solution to this problem, but they only index text. This limitation is becoming critical because the amount of digital imagery available on the Web is growing rapidly. The goal of our research is to construct search engines for images. The most important challenge is to ensure that our approach scales to very large image collections, on the scale of millions of images.

We plan to construct a search engine for images that relies exclusively on information automatically extracted from images. We are addressing the scalability challenge by designing efficient new algorithms specifically designed to produce more accurate results in search engines. For example, traditional histogram-based matching methods only use the distribution of different colors or intensity within the image, and thus contain no spatial information. We have designed a method called histogram refinement which imposes additional constraints on histogram-based matching, and allows spatial information to be taken into account when comparing images. Our initial work has concentrated on spatial coherence, but we are extending histogram refinement to exploit a wide variety of spatial information. Our methods have yielded a statistically significant improvement over conventional histogram-based matching, and are very efficient. For example, a collection of 15,000 images can be queried using histogram refinement in approximately 2 seconds on a Pentium workstation.

Daniel Huttenlocher, Associate Professor, Computer Science Department

To a large degree, images and video are cumbersome to use with today’s computing and communications infrastructure. The inability to interact with images and video based on aspects of their content, as we currently do with text, is one source of difficulty. For example, consider a system that could automatically synchronize a video recording of a lecture with the presentation materials (e.g., identify which overhead, slide or computer image is being displayed). Such a system would enable viewers to have two separate displays, one showing the presentation materials and one showing the video of the presentation. This would solve the problem that video is generally too low-resolution for the materials to be readable. More important, it would enable the materials to be used as a content-based index into a presentation, thereby allowing viewers to search, browse, or view those sections that are of interest.

Daniel Huttenlocher, Associate Professor, Computer Science Department

One of the great promises of global networked computing environments is the ability to facilitate groups working together across time and space. Within the context of university education, we have been developing and experimenting with systems that enable students to work together from around the campus and the community, and that make it easier for groups of people to share information. The systems that we have developed are currently text-based, but with the proposed infrastructure we plan to create remote collaboration and learning systems that make use of images, video and audio. We expect that adding these modalities, which for many users are more natural than text, will expand the applicability of our systems beyond the realm of engineering education where they have primarily been used to date.

CoNote is a computer supported cooperative work system designed to facilitate communication within a group via the use of shared annotations on a set of documents. The central idea underlying the system is that shared annotations provide an effective forum for groups whose work involves frequent reference to some set of documents. In our experience, the shared annotations model provides a richer electronic forum than media such as news groups, bulletin boards or mailing lists. The key difference is that the documents being annotated provide a context for group discussions, thus enabling people to find discussions on particular topics more easily. We plan to add shared annotation capability to other Web-based applications such as the video presentation browser described above.