DeCon-Net: decoupled hierarchical contrast for soccer object detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Sports Analytics

Background:

• Accurate object detection is crucial for automated soccer video analysis in broadcasting, tactics, and training.
• Existing methods struggle with feature learning, leading to 'feature collapse' where player and ball features become indistinguishable or lost.
• Dense scenes in soccer exacerbate these limitations due to insufficient feature discriminative capability.

Purpose of the Study:

• To address the 'feature collapse' phenomenon in soccer object detection.
• To enhance the discriminative capability of models in dense soccer scenes.
• To improve the accuracy of detecting players and the soccer ball.

Main Methods:

• Proposed DeCon-Net, featuring a Decoupled Feature Learning Module (DFLM) and a Hierarchical Contrastive Constraint Module (HCCM).
• DFLM utilizes dual-stream encoders for separate appearance and identity feature extraction with mutual exclusivity constraints.
• HCCM implements dynamic threshold contrastive learning for progressive feature optimization from coarse to fine granularity.

Main Results:

• DeCon-Net demonstrated significant performance improvements on the SportsMOT and SoccerNet-Tracking datasets.
• Substantial gains were observed in soccer ball detection accuracy.
• The proposed method effectively mitigates feature collapse and enhances detection in dense scenarios.

Conclusions:

• DeCon-Net successfully overcomes the limitations of feature collapse in soccer object detection.
• The novel DFLM and HCCM modules provide a robust framework for improved feature learning and contrastive constraints.
• This approach offers a significant advancement for automated soccer video analysis, particularly in challenging dense scenes and ball tracking.