Shahrokh Heidari ¹ Gibran Zazueta ² Riki Mitchell ³ David Soriano Valdez ² Mitchell Rogers ¹ Marcel Noronha ⁴ Alfonso Gastelum Strozzi ² Mengjie Zhang ⁵ Patrice Jean Delmas ^6*

• ¹ IVSLab, The University of Auckland, Auckland, New Zealand
• ² UNAM Mexico City, Mexico City, México, Mexico
• ³ Riki Consulting Ltd, Auckland, New Zealand
• ⁴ One New Zealand Warriors, Auckland, New Zealand
• ⁵ Victoria University of Wellington, Wellington, Wellington, New Zealand
• ⁶ The University of Auckland, Auckland, New Zealand

The application of Artificial Intelligence (AI) and Computer Vision (CV) in sports has generated significant interest in enhancing viewer experience through graphical overlays and predictive analytics, as well as providing valuable insights to coaches. However, more efficient methods are needed that can be applied across different sports without incurring high data annotation or model training costs. A major limitation of training deep learning models on large datasets is the significant resource requirement for reproducing results. Transfer Learning and Zero-Shot Learning (ZSL) offer promising alternatives to this approach. For example, ZSL in player reidentification (a crucial step in more complex sports behavioral analysis) involves re-identifying players in sports videos without having seen examples of those players during the training phase. This study investigates the performance of various ZSL techniques in the context of Rugby League and Netball. We focus on ZSL and player re-identification models that use feature embeddings to measure similarity between players. To support our experiments, we created two comprehensive datasets of broadcast video clips: one with nearly 35,000 frames for Rugby League and another with close to 14,000 frames for Netball, each annotated with player IDs and actions. Our approach leverages pre-trained re-identification models to extract feature embeddings for ZSL evaluation under a challenging testing environmnet. Results demonstrate that models pre-trained on sports player re-identification data outperformed those pre-trained on general person re-identification datasets. Part-based models showed particular promise in handling the challenges of dynamic sports environments, while non-part-based models struggled due to background interference.