ASVTDECTOR: A practical near duplicate video retrieval system

In this paper, we present a system, named ASVT-DECTOR, to retrieve the near duplicate videos with large variations based on an 3D structure tensor model, named ASVT series, over the local descriptors of video segments. Different from the traditional global feature-based video detection systems that incur severe information loss, ASVT model is built over the local descriptor set of each video segment, keeping the robustness of local descriptors. Meanwhile, unlike the traditional local feature-based methods that suffer from the high cost of pair-wise descriptor comparison, ASVT model describes a video segment as an 3D structure tensor that is actually a 3 x 3 matrix, obtaining high retrieval efficiency. In this demonstration, we show that, given a clip, our ASVTDETECTOR system can effectively find the near-duplicates with large variations from a large collection in real time.