playbox
playbox
Contact
Back to Posts List
4th place in the world at the SoccerNet GSR Challenge, a competition of CVPR 2025, the world's premier international conference in the field of AI and computer vision

4th place in the world at the SoccerNet GSR Challenge, a competition of CVPR 2025, the world's premier international conference in the field of AI and computer vision

Date published:2025/6/25

~Collaborative R&D with MIXI. Highly Accurate Automatic Estimation of "State Recognition" from Soccer Match Footage

(Head Office: Chiyoda-ku, Tokyo; CEO: Scott Atom; hereinafter "Playbox") and MIXI Inc. (Head Office: Shibuya-ku, Tokyo; President and Senior Executive Officer CEO: Hiroki Kimura; hereinafter "MIXI") have collaborated on the "SoccerNet Challenge (SoccerNet Challenge)," a competition at the world's most prestigious international conference in the fields of AI and computer vision, CVPR 2025. (Head Office: Shibuya-ku, Tokyo; President and Senior Executive Officer: Hiroki Kimura; hereinafter "MIXI"), participated in the SoccerNet Challenge (Game State Reconstruction) (*1) (hereinafter "SoccerNet GSR Challenge"), a competition at CVPR 2025, the world's premier international conference on AI and computer vision, and achieved the 4th place in the world.

This challenge is an advanced AI task to reconstruct the position, movement, and game state of players and the ball from soccer game footage, and is a competition for analyzing the accuracy and applicability of sports footage. Through our collaboration with MIXI, we have promoted joint research and development of video analysis algorithms and worked to build a sports analysis technology that can be used in actual operations.

CVPR (*2) is the world's premier international conference in the fields of AI and computer vision, and the largest gathering of top researchers in image recognition technology. The SoccerNet GSR Challenge, a competition held in conjunction with CVPR, has been held every year since 2021 to compete in AI technology for the automation of "Game State Reconstruction" using soccer game footage.

Since December 2024, we have been collaborating with MIXI's AI Modeling Group on research and development to build an analysis platform that utilizes data from tracking and event detection as well as AI analysis of sports videos. Both companies jointly competed in this year's SoccerNet GSR Challenge, and out of 76 teams (*3) from around the world, we were rated 4th in the world.

Outline of Research and Development

State Recognition (GSR), which ascertains the roles (field players, goalkeepers, referees, etc.) and positional information of each person from soccer match video, is a task to reconstruct the match situation in a 2D view and is a very important technology in tactical analysis and play evaluation To realize GSR, A combination of multiple AI technologies such as field and person detection, player identification (team and number), player tracking, and positional estimation is required, and each of these technologies must be improved in order to be put into practical use.

In this effort, we are using more accurate deep learning models, geometric inference (camera calibration), and domain knowledge based on soccer rules and player roles (e.g., goalkeepers are mainly located near the goal area/referees may wear uniforms of similar colors to the players ), we have succeeded in improving accuracy significantly over existing methods.

Original video courtesy of SoccerNet [Giancola et al., 2018] / Tracking results: from this effort

Future Prospects

With the presented analysis pipeline, it is now possible to obtain detailed information from soccer match videos with a high degree of accuracy, such as who belongs to which team, what is their number, what role they play, and where they are on the pitch. In particular, the "camera calibration," which estimates the position on the pitch according to camera movement, has greatly improved recognition accuracy and has been highly evaluated in international benchmarks.

The technology to achieve highly accurate state recognition (GSR) is still in its infancy, and there remains room for improvement in the future, such as increasing the accuracy of player role and team classification, and further stabilizing back number identification. In addition, the current analysis pipeline has some issues that require time-consuming processing, and we plan to increase parallelization and inference optimization to achieve faster and more efficient processing and improve its practicality. In the future, we intend to have AI deeply learn soccer-specific domain knowledge to realize a system that can understand and analyze situations more like a human being.

The knowledge and technology gained through this initiative will provide important clues for teams and players to objectively understand their own performance and improve the quality of their tactics and play. Playbox will promote the development of sports analysis technology that is more accurate and easy for anyone to use. In the future, Playbox aims to provide this technology to a wide range of sports teams, from amateurs to professionals, and contribute to the creation of new ways to enjoy and value sports.

Comments from the participating members of the SoccerNet GSR Challenge 2025

Rio Watanabe, Manager, AI Modeling Group, Dandelion Room, Development Division, MIXI Inc.

MIXI is exploring the possibilities of video analysis using computer vision technology and applying AI technology in the field of sports, mainly in the Development Division. Through our joint research with Playbox, we took on the extremely challenging task of automatically estimating the roles and positions of players from soccer match footage, and were able to achieve 4th place in the world in this competition. We will continue to utilize AI technology to develop experiences and technologies that allow people to enjoy sports more deeply.

Scott Atom, CEO and President, Playbox, Inc.

Playbox continues to take on the challenge of expanding the possibilities of sports through the use of AI technology based on the vision of "making human movement computable. In this achievement, we were able to achieve higher accuracy than last year when we used private data only with public data, and we realized that AI sports analysis technology is evolving at a tremendous pace. Playbox will continue to deliver advanced and interesting data to more people and create products and businesses that enhance the value of sports. Last but not least, we would like to express our sincere gratitude to MIXI, with whom we have collaborated, and to the many other parties involved for their support.

(*1) SoccerNet Challenge (Game State Reconstruction): An international competition on Game State Reconstruction (GSR) tasks organized by SoccerNet. GSR is a hot topic that can be widely applied to sports data analysis and tactical analysis, and is attracting a great deal of interest from the industrial world.

(*2) CVPR (The IEEE/CVF Conference on Computer Vision and Pattern Recognition): The world's premier international conference where industry-leading research results in computer vision, artificial intelligence, machine learning, and related fields are presented. Conference.

(*3) Based on the number of participants shown on the SoccerNet leaderboard.

For inquiries regarding this matter, please contact

Playbox Inc.
Public Relations
Email: info@play-box.ai
Web site: https://www.play-box.ai