July 24, 2001
Computer vision is playing a key role in the development of tools that allow the user to easily find or select audiovisual contents. Videos should be described by their semantic contents in order to use the same language as the user. The « affective » characterization of the contents of a video is given in terms of a high-level structure or by semantic descriptors that may give the user a more appropriate knowledge about it. Computer vision allows us to obtain visual descriptors that are directly related to semantics through additional high-level knowledge.
The first step in our research consists of finding out how far we can reach with different visual cues. Color and motion information seem to be the most informative ones. We are currently working on the characterization of motion using probabilistic models, particularly Gibbs/Markov random fields (GMRF). Some preliminary results will be shown. Unsupervised clustering using Kullback-Leibler divergence (KLD) as a « similarity » measure gives promising results, as the clusters obtained can be interpreted in terms of their high-level contents. A deeper analysis of the results is still to be done, as well as the application of different more complex motion models.
Juan Ma Sanchez, a PhD candidate from Barcelona is giving us a short visit this week. He will give a seminar tomorrow (Tuesday, July 24) at noon, Barus & Holey, Room 190 (new extension).
Juan Ma is presently spending his summer in NYC (working on video image analysis) @ Columbia in the lab of John Kender.
Last Updated: July 24, 2001