Nanjing University in TRECVID 2009_final

This report describes the details of our system for copy detection task in TRECVID 2009. We submitted 4 runs, 2 for video-only queries and 2 for video + audio queries. Details are listed below. Each group is with same technique but different parameters. NJU.v.balanced.1, NJU.v.nofa.1: SURF feature is used. Videos are indexed with feature points. We adopt LSH as indexing technique. NJU.m.balanced.1, NJU.m.nofa.1: SURF feature is used. Feature points are made to BOF before videos are indexed. A post process of SURF points pairing is included. For video-only queries, LSH (Locality-sensitive Hash) was built using SURF points but the approach cost too much memory. To save time and space cost, we generated a global feature for each frame using bag-of-features when dealing with video + audio queries, and ignored audio information in both query and reference media. Because it’s our first time participation, our goal is to complete whole process of TRECVID copy detection task, get experience of the key technologies in CBCD.