An Efficient Partial Video Copy Detection for a Large-scale Video Database.

Partial video copy detection (PVCD) aims to discover copy segments of query videos from a video database, which plays an important role in video copyright protection, filtering, tracking, etc. For a large-scale video database, PVCD can be divided into two stages: the first stage involves searching for video-level copies of the query video in the database, and the second stage is to further localize the copy segments within the video-level copies. Thus, two major challenges arise: (1) efficiently and effectively calculating the similarity between videos; (2) localizing mixed-duration video pairs. To address the above challenges, we propose an efficient PVCD approach for a large-scale video database, based on the Bag-of-Words (BoW) framework, which decouples video-level similarity and copy localization into cell-level. This approach consists of two modules. The first is an efficient video similarity measurement (VSM) module for the large-scale video database. VSM aggregates cell-level similarity into video-level similarity, and with a dual index, it greatly improves retrieval speed while accurately measuring spatiotemporal transformations. The second is a greedy pattern detection (GPD) module for video copy localization. GPD quickly and accurately detects similarity patterns through a greedy strategy on the similarity matrix formed by matching frames in each cell, then aggregates them into complete predicted copy segments. On the comprehensive dataset self-SVD, VSM significantly outperforms state-of-the-art methods by 7.28% in mAP, and the retrieval speed is increased by over 318 times. Additionally, for short videos at the scale of hundreds of millions, the response speed can theoretically reach seconds. On the copy localization dataset MIX, composed of mixed-duration videos, GPD also achieves the best performance.