The advent of deep learning tools has significantly
enhanced researchers’ capabilities in analyzing
spatio-temporal data. This type of data analysis
holds relevance across various domains.
Improving the ability to identify and cluster patterns
within sporting events has profound implications.
It can aid in automatic highlight detection
and is especially beneficial for coaching, particularly
in underfunded and minor leagues. While the insights
presented in this paper can be applied to numerous
team sports, our focus primarily lies on ice
hockey. In this paper, we make three significant contributions:
we introduce a simple, parametrized ice
hockey formation dataset generator facilitating the
development and benchmarking of baseline models;
we investigate the impact of noise from the
dataset on the accuracy of event classification; and
we compare the accuracies of three models: KNearest-
Neighbors, Graph Networks, and Convolutional
Neural Networks.
PDF