KDDI LABS and SRI International at TRECVID 2010: Content-Based Copy Detection

We describe our system for Content-Based Copy Detection (CBCD) task submitted to TRECVID 2010. Our system is multi-modal and integrates the results of global visual features, local visual features and audio features to produce the final run results. Each of these features is designed to take care of different aspects of the video transformations. We submitted two runs each for BALANCED as well as NOFA profile: • KDDILabs-SRI.m.balanced.1 • KDDILabs-SRI.m.balanced.2 • KDDILabs-SRI.m.nofa.1 • KDDILabs-SRI.m.nofa.2 These runs all use the same CBCD framework for each of the three modalities and differ only in the parameters for the final integration step. Our CBCD framework has made significant advances in video based copy detection. We introduce novel algorithms to obtain robust results against various transformations: dense-sampling-based global SIFT features, improved indexing methods for both global and local features and handling temporal burstiness. TRECVID 2010 evaluation results show that our system achieves good performance for both detection accuracy especially on NOFA profile and localization accuracy.