Event Detection in Ice Hockey Using Game-Aware Representations: A Dataset and Baseline Study

Abstract

Ice hockey occupies a significant cultural and economic position in Canada, yet analytical tools for understanding the game remain relatively underdeveloped. In particular, event detection is still performed manually, hindered by several challenges: the absence of publicly available, labeled ice hockey event datasets; the limited transferability of event detection methods developed for other sports due to differing event semantics; and the lack of approaches that incorporate contextual priors, here referred to as game attributes, such as player locations or team structure. Automating event detection through the integration of such game attributes has the potential to greatly enhance the analysis of broadcast footage by enabling efficient detection and review of key moments, improving player development pipelines, and supporting real-time highlight generation, especially in leagues without access to sophisticated tracking systems. This work addresses these gaps by introducing a small scale, fully annotated ice hockey event dataset, establishing baseline models for benchmarking, and comparing three distinct data representations: RGB frames, occupancy grids, and graph-based structures, some of which explicitly incorporate game attributes to improve detection performance.