Content-Based Video Copy Detection: PRISMA at TRECVID 2010

We present PRISMA’s Video Copy Detection system (P-VCD). The system is based on visual-only global descriptors, weighted combinations of distances, a pivotbased index structure, and a novel approximated search and voting algorithm for copy localization. We submitted four Runs to TRECVID 2010 CCD task: PRISMA.m.balanced.ehdNgryhst: a combination of edge histogram and gray histogram. PRISMA.m.balanced.ehdNclrhst: a combination of edge histogram and color histogram. PRISMA.m.nofa.ehdNgryhst: a combination of edge histogram and gray histogram. PRISMA.m.nofa.ehdNghT10: a combination of edge histogram and gray histogram with a different threshold. P-VCD’s results show that the combination of edge histogram and gray histogram is slightly better than edge histogram and color histogram. These results were positioned above the median, and considering just video-only Runs, were the bests positioned for Balanced and Nofa profile. These results also show that our pivot-based index enables to discard 99.9% of distance evaluations and still have good effectiveness, and that global descriptors can achieve competitive results with TRECVID transformations.