University of Marburg at TRECVID 2007: Shot Boundary Detection and High Level Feature Extraction
Markus Mühling, Ralph Ewerth, Thilo Stadelmann, Christian Zöfel, Bing Shi, Bernd Freisleben · 2007
In this paper, we summarize our results for the shot boundary and high level feature detection task at TRECVID 2007. Our shot boundary detection approach of previous TRECVID evaluations served as a basis for our experiments this year and was modified in several ways. First, we have incorporated a new metric selection for cut detection based on the evaluation of a clustering result. Second, we have tested the possibility to improve cut detection results via self-supervised learning. Third, the unsupervised approach for gradual transition detection has been supplemented with a false alarm removal method using a state-of-the art camera motion estimation approach. Regarding high-level feature detection, one focus of this year’s task was to investigate the question how well a trained system generalizes from the TRECVID 2005 news data to this year’s Sound and Vision data. However, only two institutes have submitted four runs of the related type “a ” for evaluation (three of them were submitted by us). In this paper, we present our experiments for the high-level feature task with respect to the generalization capabilities of our system trained on broadcast news videos. For this purpose, we have conducted several experiments using our system which is based on low-level features as well as on state-ofthe-art approaches for camera motion estimation, text detection, face detection and audio segmentation. In this section, the results of our participation in both tasks are presented in form of the requested structured abstract. The shot boundary detection approach and the related experimental results are presented in section 2. Our system developed for high-level feature extraction is described in section 3 along with the experimental results. Section 4 concludes the paper. The following definitions are used in this paper: # correctDetectedItems recall = (1) # Items # correctDetectedItems precision = (2) # correctDetectedItems falseAlarms 2 * recall * precision f 1 =