Large-scale Efficient and Effective Video Similarity Search

Abstract

Recently, the rich diversity of the video capture devices and the high usage of the Internet have generated a great amount of video data, which attracts the attention of researchers with respect to the development of novel effective and efficient video retrieval approaches. In this paper, we investigate the effectiveness and efficiency of the lower-bounding filter distance functions of the well-known similarity measure Earth Mover's Distance (EMD) on signature databases, including the recently introduced Independent Minimization for Signatures (IM-Sig). We conduct the experiments on a public dataset comprising various categories with visually similar videos, and another large-scale real world video dataset consisting of 350,000 near-duplicate videos. To the best of our knowledge, this is the first work investigating the effectiveness and efficiency of the lower-bounding filter distance functions on databases consisting of signatures, i.e adaptive-binned representations. The experimental evaluation indicates both high effectiveness and efficiency of the IM-Sig, outperforming the state-of-the-art techniques.