High-accuracy game state reconstruction from soccer match footage achieved through collaborative R&D with Playbox

Tokyo, Japan, June 25, 2025 — MIXI, Inc. is pleased to announce that our company placed fourth in the world in the SoccerNet Game State Reconstruction Challenge*¹ at CVPR 2025, one of the world’s leading international conferences in the fields of AI and computer vision.
For this competition, we conducted research and development through collaboration with Playbox Inc. (hereinafter, Playbox) and entered the competition as a joint team.

CVPR*² is one of the world’s leading international conferences in AI and computer vision, and the largest academic conference bringing together leading researchers in image recognition technology. Held alongside CVPR, the SoccerNet GSR Challenge has been held annually since 2021 and focuses on AI technologies for automating game state reconstruction (GSR) using soccer match footage.
MIXI’s AI Modeling Group has been working with Playbox since December 2024 on research and development aimed at building an analytics platform that uses AI analysis of sports footage, along with tracking and event detection data. The two companies entered the SoccerNet GSR Challenge together this fiscal year. In a competition requiring highly accurate analysis for technical reports, the joint team from MIXI and Playbox ranked fourth worldwide among 76 teams*³.

Overview of Research and Development

GSR reconstructs match situations in a 2D view from soccer footage, identifying each person’s role, such as outfield player, goalkeeper, or referee, and their location on the field. GSR is a key technology for tactical analysis and play evaluation. GSR relies on a combination of AI technologies, including field and player detection, player identification (team affiliation and jersey number), player tracking, and positional relationship estimation. Further improvements in each of these technologies are needed to make them viable in real-world applications.

In this initiative, we achieved a significant improvement in accuracy over existing methods by combining advanced deep learning models, geometric inference (camera calibration), and heuristic optimization based on domain knowledge of soccer rules and player roles. Examples of this domain knowledge include that referees may wear uniforms similar in color to player uniforms and goalkeepers are typically located near the goal area.

Future Outlook

The analysis pipeline presented here makes it possible to accurately identify detailed information from soccer match footage, including who each player is, which team they belong to, their number, their role, and where they are on the field. In particular, recognition accuracy has improved significantly in camera calibration, which estimates positions on the field in response to camera movement, and it has earned high marks in international benchmarks.
On the other hand, the technology required to achieve highly accurate GSR is still being developed, with room for improvement in areas such as increasing the accuracy of player role and team classification and further stabilizing jersey number identification. In addition, the current analysis pipeline is time-consuming, so we plan to improve its practical usability by advancing parallelization and inference optimization to enable faster and more efficient processing. Looking ahead, we aim to build a system that can be trained more deeply on soccer-specific domain knowledge, allowing it to understand and analyze situations in a more human-like way.

We believe that the insights and technologies gained through this initiative can also be applied to sports tactical analysis, training, and enhancing the fan experience. As we continue our research and development toward more advanced analytical technologies, we aim to create new added value in communication through our distinctive AI-driven innovation.

Comments from participants in SoccerNet GSR Challenge 2025

・Scott Atom, President and Representative Director, CEO of Playbox, Inc.
At Playbox, our vision is to make human movement computable. We take on challenges to expand the potential of sports by utilizing AI technology. We achieved higher accuracy this year using only public data than we did last year using private data, making us feel firsthand how rapidly AI sports analysis technology is advancing. Going forward, Playbox will continue to deliver advanced and engaging data to more people, and create products and businesses that enhance the value of sports. Lastly, I would like to express my heartfelt gratitude to all involved, including MIXI, for their collaboration.

・Rio Watanabe, AI Modeling Group Manager (Tanpopo Division, Development Department) at MIXI, Inc.
At MIXI, our Development Division is at the heart of exploring video analysis with computer vision, and we are working on applying AI in sports. Through joint research with Playbox, we took on the highly challenging task of automatically inferring player roles and positions from soccer match footage, and we achieved fourth-place worldwide in the competition. Going forward, we will continue developing AI technologies not only for practical use in tactical analysis on the field, but also to create richer viewing experiences.

About Playbox Inc.

Playbox, Inc. is a startup originating from the University of Tsukuba and Nagoya University. Built around technology that makes human movement computable, Playbox supports decision-making and action by people and organizations and aims to create a future where people and technology co-create. Beginning with “Playbox”, an AI sports camera that makes automated filming and editing easy, the company will continue to implement advanced video analysis and AI technologies across various fields beyond sports.
Established: November 28, 2024
Location: 1-6-16 Kanda Izumicho, Chiyoda-ku, Tokyo, Yamato Building 405
Representative: Scott Atom
Corporate website: https://www.play-box.ai/