Player detection method based on scale attention and scale equalization algorithm

Area of Science:

• Computer Vision
• Machine Learning
• Sports Analytics

Background:

• Object detection in team ball games faces challenges with scale variations, leading to missed detections for smaller players and reduced accuracy for larger ones.
• Current methods often rely heavily on dataset scale statistics, which can be unreliable and lead to performance degradation.

Purpose of the Study:

• To develop an improved object detection method for team ball game players that overcomes limitations related to bounding box scale.
• To enhance detection accuracy for players across various scales, particularly addressing underperformance with small and large bounding boxes.

Main Methods:

• A novel multi-scale attention mechanism is proposed, reducing reliance on scale statistics by using a specially created SIoU (Similar to Intersection over Union) label.
• The SIoU label explicitly represents multi-scale features, guiding the training of multi-scale attention network modules at two granularity levels.
• An integrated scale equalization algorithm within SIoU labels is employed to improve the detection of multi-scale targets in imbalanced datasets.

Main Results:

• Experiments on basketball, volleyball, and ice hockey datasets demonstrate the effectiveness of the proposed method.
• The approach achieved significant improvements in detection accuracy for players with smaller bounding boxes (up to 11%) and larger bounding boxes (up to 15%).

Conclusions:

• The proposed two-fold approach effectively addresses scale-related challenges in object detection for team ball games.
• This method offers a robust solution for improving player detection accuracy across diverse scales and game scenarios.