Recognizing a Sequence of Events from Tennis Video Clips: Addressing Timestep Identification and Subtle Class Differences

Detecting temporally precise and fine-grained events from tennis videos is important in automatic video annotation. This paper addresses the challenges of recognizing a sequence of events from tennis video clips, focusing on accurate timestep identification and distinguishing subtle class differences. We propose a novel but simple end-to-end event detection network to accurately detect and identify the key events, which can be trained on a single GPU. We demonstrate that our model outperforms the existing baselines on our fine-grained tennis event dataset. The research contributes to the development of tennis video analytics and has broader implications in other sports domains.

I. Introduction

The analysis of sports videos plays a crucial role in enhancing athletic performance, refining coaching strategies, and providing valuable insights into player dynamics. In the realm of tennis, the ability to recognize and understand a sequence of events (e.g., players hitting the ball and the ball bouncing on the court) from video clips has gained significant interest due to its potential to empower players, coaches, and analysts alike. However, achieving accurate event detection in tennis videos presents unique challenges that demand innovative solutions.