Spatiotemporal Event Spotting via 3D Heatmaps with Dynamically Shifted Gaussian Kernels

Abstract

Spatiotemporal event spotting in ball sports is essential for understanding complex game dynamics, requiring both high temporal precision and accurate spatial localization. However, most existing methods focus primarily on temporal localization, often neglecting the spatial dimensions that are crucial for tactical analysis. In this study, we propose a novel representation, 3D Heatmaps with Dynamically Shifted Gaussian Kernels, specifically designed to enable comprehensive spatiotemporal event spotting. To overcome current limitations, we introduce the Volleyball Nations League (VNL) dataset, which includes detailed annotations for eight key event types, encompassing both temporal and spatial labels. Our approach leverages a modified 3D U-Net architecture that effectively captures spatiotemporal patterns by utilizing our proposed heatmap design. Experimental results show that our method significantly outperforms state-of-the-art techniques in both temporal accuracy and spatial precision on the VNL dataset and a spatially augmented version of the SoccerNet Ball Action Spotting (BAS) dataset. These findings demonstrate the robustness and generalizability of our approach across different ball sports domains.