Siamese Neural Networks for Player Re-identification in Basketball: TwinPlay.ai Approach

Computer vision technology has revolutionized sports analytics by enabling automated analysis of athlete performance and game dynamics through sophisticated video processing algorithms. In basketball, these systems facilitate real-time player tracking and performance evaluation, transforming how coaches approach game analysis. However, player re-identification in basketball presents significant challenges due to similar player appearances, dynamic movements, and variable imaging conditions. This thesis investigates Siamese Neural Networks for basketball player re-identification, focusing on systematic testing and performance enhancement through architectural optimization and training procedure improvements. The research employs a Siamese triplet network architecture based on EfficientNet-B0, integrated with the TwinPlay.ai basketball analytics platform. A comprehensive experimental framework evaluates triplet loss margins ranging from 1 to 5000, analyzing their impact on embedding space geometry and re-identification performance. The methodology incorporates enhanced data validation, robust preprocessing with image resizing, and comprehensive metric collection including embedding statistics and distance distributions. Key findings demonstrate that triplet loss margin selection critically affects embedding space structure and requires systematic optimization rather than arbitrary selection. The investigation reveals complex relationships between margin values and embedding space geometry, where larger margins generally improve separation between positive and negative pairs, but optimal performance depends on finding the appropriate balance for the specific application context. Systematic margin analysis shows that while increasing margins can enhance inter-class separation, the relationship is non-linear and dataset-dependent, emphasizing the importance of empirical evaluation rather than assuming larger values are universally superior. Peak separation analysis demonstrates substantial improvements in embedding space organization when margins are properly tuned, achieving optimal separations in both anchor-positive and anchor-negative distributions. Technical contributions include enhanced data validation mechanisms, comprehensive distance distribution analysis with histogram-based peak detection, improved training metrics collection, and integration of threshold-based accuracy computation functions. The evaluation framework incorporates multiple performance metrics, demonstrating superior embedding space organization compared to baseline implementations. The research provides practical guidelines for optimal hyperparameter selection in basketball player re-identification applications, with direct applications in sports analytics for improved player tracking, performance analysis, and automated game statistics generation. Results indicate that systematic optimization of Siamese architectures through careful margin selection significantly enhances basketball player re-identification performance in real-world sports analytics environments.