Radu S. Jasinschi, Philips Research, Briarcliff Manor, U.S.A.
Title: The Extraction, Composition, and Description of 3-D Camera Velocity and Depth: Methods and Applications
Abstract. The use of 3-D natural visual scene information has been the topic of research in the last 30 years in computer vision, image processing, computer graphics, and pattern recognition. With the development of fast computers and communiction systems, the use of 3-D information, being it natural or synthetic, is becoming increasingly important for applications such as virtual environments and video query. In this talk I present some basic tools that I co-developed for the extraction, composition, and description of 3-D natural visual information. I discuss a robust method for the automatic extraction of 3-D camera velocity and static 3-D scene depth. This method does not require any knowledge about 3-D scene content information. Next, I describe work on layered sprites. This work extends mosaicing from a single mosaic to a layered set of mosaics by incorporating 3-D scene information into the mosaicing process. We also describe how to generalize this process for the generation of 3-D sprites. Finally, I discuss a video query system based on a MPEG-7 descriptor, which uses the 3-D camera and scene parameters defined above. This descriptor was incorporated into the MPEG-7 experimentation model, and it is now part of the working draft. It is used to query for video shots for which the camera exhibits a similar motion type based on the user's requirements. We show experimental results using these tools for the generation of layered sprites and for video query. We conclude this talk by briefly discussing future developments of this work for virtual environments and for the MPEG-7 standardization effort.