Shot Boundary Detection in Soccer

Abstract

The production of multimedia videos has surged thanks to affordable video equipment, large storage devices, and user-friendly editing tools. As video content proliferates, efficient analysis becomes increasingly critical to reducing production costs and manual labor. A fundamental step in video analysis is shot boundary detection (SBD), which segments videos into discrete units called shots. However, detecting shot boundaries poses significant challenges due to the diversity of transition types and frequent domain-specific variations. For example, a method that performs well on TV news may fail when applied to sports videos. This study compares popular SBD algorithms within the soccer domain, revealing important insights. It was found that pixel-based approaches outperform the histogram-based approach in terms of precision, while deep learning models, such as TransNetV2, struggle with certain types of transitions, particularly logo transitions. Building on these findings, we developed a hybrid algorithm that combines pixel and deep learning approaches to improve performance. The proposed algorithm achieved significant performance improvements, with an average F1 score of 0.94 in the La Liga tournament and 0.88 in five other European tournaments. While the algorithm is expected to perform well in various soccer events such as penalties, fouls, and different tournaments due to the similar structure of transitions and video content, variations in camera angles, equipment quality, and broadcast production styles may impact its performance. Key contributions of this work include the creation of a specialized soccer dataset, a comprehensive comparison of widely used SBD algorithms, and the development of an improved SBD algorithm tailored specifically for soccer videos. Future research will aim to expand the dataset, enhance the algorithm's capability to detect more complex transitions, and optimize its performance for real-time applications.