Toward Automatically Labeling Situations in Soccer.

Area of Science:

• Computer Science
• Machine Learning
• Sports Analytics

Background:

• Automatic annotation of soccer game situations is challenging due to the scarcity and cost of expert-labeled data.
• Traditional supervised learning methods struggle with limited annotated data, hindering performance.

Purpose of the Study:

• To develop an effective method for automatic soccer event annotation.
• To overcome the limitations of fully supervised approaches by leveraging unlabeled data.

Main Methods:

• A two-part approach was employed: learning feature representations with variational autoencoders (VAEs) on large unlabeled datasets.
• A large-margin classifier was trained on these learned features using only a few manually annotated examples.

Main Results:

• Four VAE architectures were proposed and evaluated.
• High predictive accuracies, exceeding 90% AUC, were achieved for detecting corner kicks, crosses, and counterattacks.
• Performance was robust across different tasks.

Conclusions:

• The proposed hybrid approach effectively addresses the challenge of limited labeled data in soccer event annotation.
• Variational autoencoders are powerful tools for learning meaningful representations from unlabeled sports data.
• This method offers a scalable and accurate solution for sports analytics.