Speaker

Abstract:

Sports provide a challenging setting for computer vision and machine learning under limited annotated data, because human bodies interact dynamically, camera conditions vary widely, and labels such as player identity, pose, events, tactical roles, and decision quality are costly and domain-dependent. This talk presents our work on sports analytics with a focus on body, skeleton, and multi-person motion data. I will introduce image-based pipelines that convert broadcast and fixed-camera videos into structured representations, including trajectories, pose/skeleton data, events, and calibrated spatial configurations. I will then discuss how these representations support subsequent modeling of individual and collective behavior, such as trajectory prediction, counterfactual evaluation, and multi-agent decision modeling. Finally, I will describe our efforts to address the annotation bottleneck through sport-specific datasets, benchmarks, open-source tools, and community challenges, and discuss future directions.

Bio:

Keisuke Fujii is an Associate Professor at the Graduate School of Informatics, Nagoya University, Japan. After receiving his Ph.D. from Kyoto University, he held a postdoctoral position at Nagoya University and worked as a research scientist at RIKEN before assuming his current position. Since April 2026, he has also been cross-appointed as a Founding Lead at CyberAgent, Inc. His research interests include computer vision, multi-agent modeling, and the integration of domain knowledge and machine learning for analyzing multi-body time-series data, including sports, animal groups, and other moving objects. In particular, he aims to develop and democratize AI in these fields and founded OpenSTARLab with his colleagues. His open-access book, Machine Learning in Sports, was published in April 2025.