Augmenting Pass Prediction via Imitation Learning in Soccer Simulations

Takeshi Kaneko, Rei Kawakami, Takeshi Naemura, Nakamasa Inoue; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2024, pp. 3194-3203

Pass analysis in soccer is essential for predicting players' actions and optimizing team strategies. Existing pass prediction methods involve supervised learning which requires costly annotations about who passes where and when. We propose the use of additional synthetic data generated by a soccer simulator to overcome this challenge. Specifically we employ imitation learning to train a policy network that mimics player behavior patterns using the data intended for prediction. This policy network along with the simulator is used to generate synthetic data. The generated synthetic data is then combined with real-world data to learn pass prediction by an existing model that utilizes both trajectory and video data. Experiments confirm that our approach improves the top-1 prediction accuracy of the intended pass receiver by 3.72% compared to an existing state-of-the-art method.