Fast and Label-Free Soccer Scene Retrieval Using Deep Learning

Abstract

With the advancements in tracking technology, an abundance of player and ball trajectory data in soccer is now being generated, leading to a growing interest in high-speed scene retrieval. However, traditional retrieval methods rely on annotated scene labels, which can be time-consuming and costly to produce. In this study, we propose a deep learning approach for fast and label-free retrieval of soccer trajectory data. The proposed method utilizes a deep learning architecture to represent the similarity between plays from trajectory data. We conduct experiments on a large set of tracking data and demonstrate that our approach outperforms traditional geometric similarity retrieval methods."