数字内容技术与服务研究中心知识库

At present, research on segment-level video copy detection algorithms mainly focuses on end-to-end optimization from key frame selection and feature extraction to similarity pattern detection, causing the deployment of such algorithms to be difficult and expensive, and ignoring specific research on optimizing detectors for similarity pattern detection. To address the above issues, we propose the segment-level Video Copy Detection Transformer (VCDT), a transformer-based detector designed for similarity pattern detection. Its main novelty can be summarized by two points: (1) An anchor training strategy that allows the model to use the positional prior information in the anchor boxes to make predictions more precisely, (2) A query adaptation module to fine-tune the anchor boxes dynamically. Our experiments show that, without bells and whistles, VCDT achieves state-of-the-art performance while showing an impressive convergence speed.