Video Phylogeny: Recovering near-duplicate video relationships

Abstract

Abstract-To keep pace with the increasing popularity of image and video sharing services, several research groups have focused on the development of systems to identify similar copies images and videos on the internet. Although these techniques allow us to identify the set of near-duplicates of a document, they do not give any information about the structure of generation of the near-duplicates. In this paper, we are interested in the history of the transformations that generated a given a set of near-duplicate videos. Given a set of near-duplicate videos, our objective is to identify their causal/phylogenetic relationships. Solutions to this problem have several applications such as in security, forensics, copyright enforcement, and news tracking services.

Video phylogeny tree reconstruction using aging measures

“…To evaluate the final effect on VPT reconstruction, we make use of the standard graph-matching metrics used in other phylogeny works [6], [11], namely: Root, which is the percentage of correctly reconstructed tree roots (i.e., the originating node); Edges, which is the percentage of correctly reconstructed directional edges; Leaves, which is the percentage of correctly reconstructed leaves, i.e., the furthest nodes in a tree; Ancestry, which measures the percentage of correctly identified ancestral relationships among videos. All these metrics assume values in [0, 1], where 0 is the worst result and 1 means perfect reconstruction.…”

“…Solving the problem of VPT reconstruction means finding the genealogical relationships between ND videos in a pool in order to infer which sequence generated another one. In order to solve this problem, state-of-the-art solutions are inspired by works originally proposed for still images [9], [10], and basically follow a common pipeline [6], [11].…”

“…First, video phylogenetic strategies start building a complete relational graph where each node corresponds to a different video sequence, and edge labels denote dissimilarity (or alternatively similarity) relations between nodes [6], [11]. Specifically, given a pair of ND video sequences S i and S j , dissimilarity is defined as…”

“…As a matter of fact, recent researches have been focusing the analysis on the relations between different versions of the same content [6], [7], [8]. The underlying idea is that multimedia contents evolve like DNA sequences of organisms mutate in biology.…”

“…analysis permits reconstructing it by analyzing similarities between the nucleotides sequences of different organisms. Similarly, multimedia phylogeny solutions aim at building a complete relational graph, in which edge labels model the similarity/dissimilarity between every pair of ND images or videos [9], [6]. Then, the underlying PT is estimated by means of graph optimization strategies that identify which dependency relations among the different contents are the most plausible.…”