Capsule-Based Regression Tracking via Background Inpainting.

Area of Science:

• Computer Vision
• Machine Learning
• Object Tracking

Background:

• Regression trackers often struggle with extreme target-background data imbalance.
• Informative background cues are crucial for accurate object tracking.
• Existing methods often overlook the potential of background information.

Purpose of the Study:

• To propose a novel capsule-based approach for regression tracking, named CapsuleBI.
• To leverage informative background cues as primary information and target cues as supplementary.
• To enhance object tracking accuracy by addressing the target-background data imbalance.

Main Methods:

• Developed a capsule-based approach (CapsuleBI) integrating a background inpainting network and a target-aware network.
• Employed a global-guided feature construction module to enhance local features with global scene information.
• Utilized a novel background-target routing algorithm for precise target localization using multi-video relationships.

Main Results:

• The proposed CapsuleBI tracker demonstrates superior performance compared to state-of-the-art methods.
• Experimental results validate the effectiveness of leveraging background cues in regression tracking.
• The method successfully handles extreme target-background data imbalance.

Conclusions:

• CapsuleBI offers a promising new direction for regression tracking by prioritizing background information.
• The capsule-based architecture effectively models object relationships and enhances tracking precision.
• This approach significantly improves tracking accuracy, especially in challenging scenarios with data imbalance.