Sztaki @ Trecvid 2009 *

We summarize our fully automatic approach to the TRECVID 2009 Search task. Our submissions summarized in Table 1 use linear combinations of the following basic techniques. • text ASR text retrieved by the Dutch translation of selected topic terms. • image Similarity of representative frames of shots. • face Face detector output for topics involving people. • feature Total weight of high level feature classifiers considered relevant by text based similarity to the topic. We used the publicly available feature predictions. • motion Motion information extracted from videos where relevant to topic. • wide A variation of text with wider shot neighborhood considered relevant. • lattice Text retrieval based on ASR lattices where available. The combination of feature and face together contributed most to the performance of the system. In this experiment the use of lattices, although they were available only for part of the shots, did not improve over the most probable ASR output. The best ASR text based run is text + wide, a combination where more distant shots also receive partial score for a matching speech. We notice that the plain linear combination of all scores deteriorated performance. In the paper we measure independent performance of the methods and observe that feature alone would have outperformed all of our runs. An improved combination includes wide with lower weight.