Partial-copy detection of non-simulated videos using learning at decision level

Abstract

There is a renewed tendency to improve video copy detection tasks due to the involved challenges in non-simulated applications. In an adverse real-world scenario, the volume of data to process as well as the variety of transformations to which a video is exposed increases continuously. Moreover, the interest in detecting not only long videos but also short partial copies increments the difficulties in copy detection methods. Therefore, we propose a practical copy detection method able to cope with partial-copies and useful in applications where real-time processing is required. To accomplish the desirable characteristics of high precision, fast processing and scalability, we use low-cost global descriptors in combination with a decision strategy adapted from a reinforcement learning technique. Our evaluation results are satisfactory to detect short segments of at least 2-seconds length under a non-simulated and severely transformed video dataset.