Effective Content-Based Near-Duplicate Video Detection
Merih Seran Uysal, Christian Beecks, Daniel Sabinasz, Thomas Seidl · 2015
In this paper, we investigate the effectiveness of various distance-based similarity measures, including the well-known similarity measure Earth Mover's Distance (EMD), and the recently introduced Independent Minimization for Signatures (IM-Sig) on large-scale signature databases. We conduct experiments on eight different databases comprising near-duplicate video clips generated via various edit tasks. To the best of our knowledge, this is the first work investigating the effectiveness of the IM-Sig and the remainder of distance-based similarity measures on databases comprising near-duplicate video signatures (adaptive-binned representations) generated via numerous video edit tasks. The experimental evaluation points out that the IM-Sig shows very accurate results, either exhibiting the same effective results as the EMD or lying directly beyond the results of the EMD, and outperforming the remainder of the state of the art.