Heterogeneous Transfer Learning in Sports: Human Action Recognition for Gender and Outcome Prediction

Abstract

This research explores the application of heterogeneous transfer learning to achieve dual objectives: gender classification and ball-on-goal position prediction (BoGP) in soccer, by analyzing players' physical actions during free kicks. Leveraging a curated dataset of soccer players executing free kicks with manual temporal segmentation, we applied pre-trained Human Action Recognition (HAR) models from the Kinetics-400 dataset. These models were adapted for our specific tasks using transfer learning techniques, minimizing the need for extensive domain-specific data. Eleven HAR backbones were evaluated for their effectiveness in both tasks. The gender classification model achieved an accuracy of 75.4%, while the BoGP model demonstrated 69.1% accuracy in predicting the ball's direction (left or right). Additionally, we examined each HAR backbone's overall performance on gender classification and BoGP prediction, revealing significant insights into the interplay between these tasks. This study highlights the versatility and robustness of HAR models in heterogeneous transfer learning.