Eurecom at video-TREC 2004: feature extraction task

Based on the results of last year worldwide submissions to the feature extraction task, we decide to introduce more features to describe the content of shots. In particular, text and motion features are added to existing visual features. The text is by definition a semantic feature, thus it has its importance in the feature extraction task. The motion is necessary to analyze specific features such as airplane takeoff. Moreover, to take advantage of the progress of classification systems, support vector machines are used to extract semantic features from low-level features. Finally, genetic algorithms are employed to fuse data from the various classifiers and modalities.