Reproduce “Predicting Penalty Kick Direction Using Multi-Modal Deep Learning with Pose-Guided Attention” on OpenResearch

About this paper

Pasindu Ranasinghe, Pamudu Ranasinghe

Researchers developed a multi-modal deep learning framework that predicts penalty kick direction (left, middle, right) with 89.38% accuracy by integrating both visual context and player 2D pose keypoints. The system processes video segments in 22 milliseconds, demonstrating suitability for real-time analysis and goalkeeper training.

Key insights

• The fusion of both visual scene context (RGB frames) and dynamic player pose information (2D keypoints) substantially increases prediction accuracy compared to using either modality alone.
• A pose-guided spatial attention mechanism allows the model to selectively focus on crucial biomechanical cues and contextual elements, improving prediction accuracy and interpretability.
• Defining input sequences dynamically using a normalized foot-to-ball distance threshold yields more predictive information, with closer proximity to ball contact resulting in higher accuracy.

Summary by alphaXiv.