BJTU TRECVID 2007 Video Search

In this paper, we describe our experiments of search task for TRECVID 2007. This year we participated in the automatic video search subtask, and submitted six runs with different combination of approaches to NIST. Using the text only based search engine used in last year, the run F_A_1_JTU_FA_1_1 provides a baseline search result list. In order to bring up true relevant results, a multi-view based reranking method is employed for reordering the search results derived from run F_A_1_JTU_FA_1_1. Specifically, initial search results, which are represented by multiple distinct feature views, are first divided into several clusters on individual feature views. According to their relevance to query intention, clusters on each feature view are mapped into predefined ranks. With these ranked clusters on all feature views, a strikingly new Cross-Reference (CR) method are employed to fuse them into a unified result ranking. The following five runs test the effect on reranking performance of different combination of clustering methods and fusion strategies. F_A_1_JTU_FA_1_2: NCut Clustering + Late-Fusion. F_A_1_JTU_FA_1_3: NCut Clustering + Single View A. F_A_1_JTU_FA_1_4: NCut Clustering + Single View B. F_A_1_JTU_FA_1_5: NCut Clustering + Bi-Fusion. F_A_1_JTU_FA_1_6: NCut Clustering + Single View A+Single View B.