Shanghai Jiao Tong University and Alibaba collaborated to develop an AI football match analysis system called MatchVision, which can watch matches, identify key tactical moves and generate comments similar to human commentators. The system is trained on the massive "SoccerReplay-1988" data set, which contains more than 3,300 hours of footage from nearly 2,000 top league matches. MatchVision can identify 24 different game events with an accuracy rate of 84%, and is superior to existing systems in foul judgment and other aspects. More importantly, the research team plans to make the data set and model open source to promote the further development of AI in the sports field.
The technology is developed based on a large dataset called "SoccerReplay-1988", which contains nearly 2,000 complete matches from Europe's top leagues and the UEFA Champions League from 2014 to 2024, totaling more than 3,300 hours of Game clips. There are an average of 76 commentary segments per contest.
This table shows how MatchVision identifies key moments in a match and generates commentary for each scenario.
MatchVision is designed as an all-in-one system capable of multitasking, including tracking match events and generating organic commentary. The system can identify 24 different types of match events, such as goals, fouls and tactical actions. When analyzing fouls, it uses multi-angle camera footage to determine the type and severity of the foul.
Test data shows that MatchVision has an accuracy of 84% in identifying game events. It not only performs well in event recognition, but also outperforms existing systems in generating comments and calling fouls. The research team plans to open source the data set and model and publish them on GitHub for use by more researchers and developers.
Interestingly, the study found that AI and human commentators focused differently on the game. AI focuses more on technical details and tactics, while human commentators focus more on the emotional flow and backstory of the game.
The side-by-side example compares how a human commentator (GT) and an AI (ours) describe three key match moments - a controversial yellow card, a corner kick sequence and a goal play.
The researchers showed a comparison of AI and human commentary on specific scenarios such as yellow cards, corners, goals and goalkeeper saves, highlighting the different ways in which the two describe key moments in a game.
In the future, the application of MatchVision may not be limited to game commentary, but can also automatically produce game highlights, and even assist referees in making more accurate penalties, based on existing AI technologies such as offside detection.
This research marks a new era in sports analytics and AI applications, bringing a new viewing experience to football fans and professionals.
The emergence of MatchVision indicates that the application of artificial intelligence technology in the sports field will be more extensive and in-depth. In the future, more sports analysis tools based on AI may appear, bringing us a more exciting and in-depth sports event experience. This technology not only revolutionizes the way of commentating football matches, but also provides a new direction for the future development of sports technology.