Stroke Classification in Table Tennis as a Multi-Label Classification Task with Two Labels Per Stroke.

Area of Science:

• Sports Science
• Computer Vision
• Machine Learning

Background:

• Table tennis strokes are crucial game elements, necessitating accurate classification from gameplay data.
• Classifying table tennis strokes is challenging due to high visual similarity between different actions.
• Existing action recognition models struggle with the nuances of table tennis stroke identification.

Purpose of the Study:

• To develop a novel multi-label stroke classification method for table tennis.
• To improve the accuracy of table tennis stroke recognition by assigning multiple labels per stroke.
• To evaluate the effectiveness of multi-labeling and different input modalities (video, 3D coordinates).

Main Methods:

• Proposed a multi-label classification approach, assigning player posture and ball rotation/velocity labels to each stroke.
• Modified action recognition models to incorporate multiple outputs for enhanced stroke classification.
• Compared performance using video data versus 3D joint coordinates as input.

Main Results:

• Multi-label classification improved accuracy by up to 8.6% on validation data and 18.1% on test data compared to single-label methods.
• Utilizing 3D joint coordinates as input yielded higher accuracy improvements (17.1% on validation, 5.4% on test) than video data alone.
• The multi-output approach reduced classification difficulty and boosted overall accuracy.

Conclusions:

• Multi-label classification is a more effective strategy for recognizing similar table tennis strokes.
• 3D joint coordinates are a valuable input modality for improving table tennis action recognition accuracy.
• The proposed method offers a significant advancement in analyzing and classifying table tennis techniques.