Universität Karlsruhe (TH) at TRECVID 2008

In this paper, we present the system developed by the Interactive Systems Labs at Universität Karlsruhe for the TRECVID 2008 evaluation. It is the second time that we participate in the TRECVID evaluation this year. Last year, the main goal of our first participation was to develop a common software framework for multimedia processing and to build baseline systems for the shot boundary detection and high level feature (HLF) extraction tasks. The evaluation task for shot boundary detection was dropped this year, hence we focused our work on HLF extraction only. The color, texture and shape features were used as low-level features as before, while textual feature was not used. We used support vector machine (SVM) classifier for low-level feature classification. The parameters for SVMs were optimized in a grid-search scheme. A simple weighted sum fusion was applied to extract the high-level features, where the weights were calculated for each category and each low-level feature. More training data was used and evaluation results were improved compared to the last year's evaluation results.