A structure-based video representation for web video categorization

In this paper, we propose a video representation that is motivated by the problem of categorizing large web video collections. The representation focuses on capturing the properties of the temporal structure of a video and deploys low-level image features derived from the self-similarity matrix of the video. The bias of the representation towards the temporal structure is based on our hypothesis that this aspect of web video content is among those that are least affected by the enormous web content diversity and that it therefore could provide a reasonable base for comparing videos independent of irrelevant content variations. Although the validation reported in this paper was only a preliminary one, the results already indicate the potential of the proposed video representation to steer towards grouping web videos into meaningful clusters.