Content-based Video Copy Detection using Binary Object Fingerprints

Content-based video copy detection in large-scale databases is still an open issue. One of the main reasons is that geometrical attack can easily surpass the global features of the frame while local features are inefficient in terms of compact representation and computational complexity. In this paper, we propose to use binary object fingerprints to represent video frame for improving the robustness of the video copy detection system. It is because salient object can be robustly detected using advanced convolutional neural network (CNN) based object detector. We proposed to use the well-known RetinaNet for generating object regions from the input frame and then these regions are used to generate binary fingerprints for fast copy detection in the database. This approach can maintain compact representation of video frame and high searching speed by binary fingerprint searching scheme. Experimental results show that the proposed approach can achieve about 10% higher recall rate with only sacrificing 1% prediction rate on VCDB dataset.