A probabilistic framework for fusing frame-based searches within a video copy detection system

In the last few years, content-based video copy detection became an important and key tool for solving the tricky problem of video copyright protection. This problem has been heightened with the development of web video exchange platforms. In general, content-based video copy detection relies on the description of the visual content of the video. The video is segmented and a selected subset of frames, called keyframes, is described. Searching for a video is then performed by a set of similarity searches based on keyframes. These searches provide partial results that have to be integrated and fused. In this paper, we focus on this particular and crucial step. The objective is to properly fuse together partial results, to take into account the temporal coherence of the video and to be efficient (i.e. rapid). The solution we propose is based on a probabilistic framework that models the different parameters and inputs of this step and enables to deal with the temporal consistency. It also makes the process more reliable, as imprecision tolerated during the keyframe-based similarity searches has no impact on the overall accuracy. This particularly allows the speeding up of the detection process.