Kobe University at TRECVID 2009 Search Task
Kimiaki Shirahama, Chieri Sugihara, Yuta Matsuoka, Kana Matsumura, Kuniaki Uehara · 2009
In TRECVID 2009 search task, we have developed a method which defines any interesting topic from examples provided by a user, especially, positive and negative examples. Specifically, considering a large variation of features in a topic, we use “rough set theory ” which defines the topic as a union of subsets. In each subset, some positive examples can be correctly distinguished from all negative examples. Based on such subsets, we can collectively retrieve shots which show the same topic but contain significantly different features. For our method, it is crucial what kind of examples are used. To examine the influence of examples on the retrieval performance, we submitted the following three runs: 1. M A N cs24 kobe1 1: In this run, a user manually selects positive and negative examples for each topic. 2. M A N cs24 kobe2 2: It is difficult for the user to select effective negative examples for defining a topic, since a huge number of shots can be negative examples. So, in this run, we use “partially supervised learning ” which defines the topic only from positive examples, by selecting negative examples from unlabeled examples (i.e. shots except for positive examples). 3. I A N cs24 kobeS 3: In this supplemental run, from the result of M A N cs24 kobe1 1, the user selects additional positive and negative examples. Note that due to the slow search speed, this run violates the maximum time limit. From evaluation results, we find that our non-interactive methods M A N cs24 kobe1 1 and M A N cs24 kobe2 2 can achieve comparable performances to medians of interactive runs. Also, I A N cs24 kobeS 3 indicates that the performance of our method can be significantly improved by using a large number of examples. 1.