Label-reconstruction-based pseudo-subscore learning for action quality assessment in sporting events

Area of Science:

• Computer Science
• Artificial Intelligence
• Sports Science

Background:

• Existing action quality assessment (AQA) methods lack detailed feedback by only providing overall scores.
• Current AQA datasets do not include labels for substage quality assessment, limiting granular analysis.
• There is a need for AQA methods that can provide detailed, substage-specific feedback for athletes.

Purpose of the Study:

• To develop a novel method for substage quality assessment in action quality assessment (AQA).
• To address the limitations of existing AQA methods by enabling detailed feedback on movement substages.
• To create a new dataset labeling approach for improved AQA.

Main Methods:

• Proposed a label-reconstruction-based pseudo-subscore learning (PSL) method for AQA.
• Utilized overall action scores as both quality labels and training features.
• Developed a label-reconstruction algorithm to generate pseudo-subscore labels for training data.
• Fine-tuned a multi-substage AQA model using pseudo-subscore and overall score labels.

Main Results:

• The proposed PSL method successfully generated pseudo-subscore labels for training.
• The multi-substage AQA model accurately predicted action quality scores for each substage and the overall action.
• Ablation experiments confirmed the effectiveness of individual modules within the proposed approach.
• The method achieved state-of-the-art performance in action quality assessment.

Conclusions:

• The developed label-reconstruction-based pseudo-subscore learning (PSL) method effectively enables substage quality assessment in AQA.
• This approach provides detailed feedback, overcoming limitations of existing methods and datasets.
• The proposed method achieves state-of-the-art results, advancing the field of AQA in sports.