A Novel Video Matching Framework for Copy Detection †

The fixed-length sliding window approach is widely used in video search applications. This approach computes the similarity between a given query and each windowed sequence, and determines whether their similarity exceeds a predefined threshold. However, using the fixed window length and the predefined threshold is inflexible for detecting video copies that might have been derived by applying a variety of video transformations. To address this problem, we present a novel matching scheme, which estimates an appropriate window length and threshold for each matching by inferring the context of the query and the current window. Moreover, to make the matching more efficient, we adopt a coarse-to-fine strategy which utilizes the compact min-hashing signature for fast filtering and robust spatio-temporal analysis for detailed verification. Substantial experiments show that the proposed video matching framework yields the robust accuracy and efficient computation for handling every type of video transformation.