Kobe University at TRECVID 2011 Semantic Indexing and Multimedia Event Detection.
Kimiaki Shirahama, Lin Yanpeng, Kuniaki Uehara · 2011
Abstract—This paper describes our methods and experimental results for TRECVID 2011 SIN and MED tasks. For SIN task, we submitted the run L A cs24 kobe sin 1 that addresses the following two problems: The first one is an expensive computation cost for constructing an SVM with a large number of examples. To ensure the detection accuracy and speed for each concept, we developed a method that selects a small number of negative examples similar to positive examples. Such negative examples are useful for characterizing the decision boundary between shots where the concept is present and shots where it is absent. The second problem is a large variety of shots where the concept is present, due to varied camera techniques and setting. Only using a single classifier is insufficient for covering such a large variety of shots. Hence, we used rough set theory to extract multiple classification rules, that characterize different subsets of positive examples. Although the evaluation result of L A cs24 kobe sin 1 was not very good, we found that one main reason is overfitting of classification rules extracted by RST. For MED task, we aim to examine the applicability of virtual reality techniques to event detection. It is laborious and timeconsuming to collect a sufficient number of positive examples for an event. To overcome this, we create virtual examples using virtual reality techniques. We developed a method that creates virtual examples by synthesizing user’s gesture, 3D object and background images. In order to evaluate the effectiveness of this approach, we submitted two runs, c real (cs24kobe MED11 MED11TEST MEDPart SemiAutoEAG c-real) and p virtual (cs24-kobe MED11 MED11TEST MEDPart SemiAutoEAG p-virtual). c real only uses positive examples that are selected from training videos. On the other hand, p virtual uses virtual examples in addition to positive examples used in c real. Detection error tradeoff curves of c real and p virtual indicate the effectiveness of virtual examples. I.