TRECVID 2007 High Level Feature Extraction experiments at JOANNEUM RESEARCH

This paper describes our experiments for the high level feature extraction task in TRECVid 2007. We submitted the following five runs: • A jr1 1: Baseline run using early fusion of all input features. • A jr1 2: Classic early feature fusion and concept correlation. • A jr1 3: Classic late feature fusion. • A jr1 4: Late feature fusion and concept correlation. • A jr1 5: Early fusion of heuristically defined feature combinations. The experiments were designed to study both, the performance of various content-based features in connection with classic early and late feature fusion, the influence of manually (heuristically) selecting input feature combinations and the application of concept correlation. Our submission made use of support vector machines based on a variety of image and video features. The results of the experiments show that four out of five runs achieved a performance above the TRECVid median, including a run with 18 out of 20 evaluated high level features equal or above the median compared with inferred average precision. The mean inferred average precision of our baseline run is 0.056. Early fusion performed slightly better than late fusion on average, although the latter produced more scores above the TRECVid median. The experiment on concept correlation generally impaired the performance and outscored the baseline only for a few features. Heuristic low-level feature combinations displayed a rather poor performance. We assume that the good baseline is due to the effective grounding of a variety of low-level visual features and the generalization capability of the SVM framework with high-dimensional feature spaces.