ARUDet: active retrieval and uncertainty-aware detection for sports video object detection

Area of Science:

• Computer Vision
• Machine Learning
• Sports Analytics

Background:

• Video object detection in sports is hindered by high-speed motion, occlusion, and deformation.
• Current methods use uniform feature aggregation, ignoring region-specific degradation and lacking targeted restoration.
• Deterministic regression methods fail to address the probabilistic nature of object boundaries in degraded video.

Purpose of the Study:

• To develop an advanced video object detection system for sports scenarios.
• To address limitations in feature aggregation and localization accuracy in existing methods.
• To improve the reliability and robustness of object detection under visual degradation.

Main Methods:

• Proposed ARUDet (Active Retrieval Uncertainty-aware Detector) with an Active Temporal Retrieval Module (ATRM) and Uncertainty Rectified Regression Head (URH).
• ATRM identifies degradation types and retrieves optimal historical features for on-demand restoration.
• URH models geometric uncertainty using a Probabilistic Boundary Projector (PBP) and a Lower Bound Optimizer (LBO).

Main Results:

• ARUDet demonstrated significant performance improvements on multiple sports datasets.
• The Active Temporal Retrieval Module effectively restored low-quality regions using complementary information.
• The Uncertainty Rectified Regression Head improved localization quality by addressing geometric ambiguity.

Conclusions:

• ARUDet effectively overcomes challenges in sports video object detection.
• The proposed active retrieval and uncertainty-aware regression strategies enhance detection accuracy and reliability.
• The method offers a promising solution for robust object detection in complex dynamic environments.