Vasyl Chomko

March 21st, 2025 – 11:00am-11:30am, EC4-2101A

Modern sports analysis depends on data-driven models for tasks such as player tracking, event detection, and 3D pose estimation. Hockey analysis, however, is constrained by the limited availability of labeled datasets. One example is hockey stick segmentation, where identifying sticks in real-world footage is challenging due to motion blur, occlusion, and limited annotated data. This work introduces Synthetic Local Data Augmentation (SLoDA), a method designed to improve segmentation models by applying targeted transformations—such as scaling, rotation, and motion blur—directly to hockey sticks in game footage. Unlike traditional augmentation methods that modify entire images, SLoDA preserves object context, improving model robustness under motion blur and occlusions. Experimental results show that SLoDA significantly improves segmentation accuracy compared to existing augmentation techniques.

Beyond segmentation, synthetic data generation can be expanded to other areas of hockey analysis, addressing the need for better datasets in multi-view data capturing, camera position estimation, puck movement analysis, game strategy analysis, player action recognition, and out-of-broadcast-view game forecasting. To support this, we are developing a controllable Unity-based simulation that provides accurate game state reconstruction and synthetic data extraction for analysis. This system allows for adjustable perspectives, ground-truth object tracking, and full scene awareness. Instead of relying on manually labeled footage alone, these approaches offer a scalable way to generate structured, labeled data, improving hockey analysis models.