DLUT: Decoupled Learning-Based Unsupervised Tracker.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised learning is vital for object tracking, requiring accurate classification and regression.
• Current unsupervised trackers often suffer from performance limitations due to high coupling between classification and regression branches via shared cross-correlation modules.

Purpose of the Study:

• To propose a novel Decoupled Learning-based Unsupervised Tracker (DLUT) that addresses the limitations of coupled branches in existing unsupervised trackers.
• To enhance the performance of unsupervised object tracking by decoupling the learning pipelines and optimizing feature extraction.

Main Methods:

• Implemented a decoupled learning strategy to separate training pipelines for different branches, enabling independent feature exploration.
• Designed adaptive decoupling-correlation modules tailored to each branch for generating more discriminative feature response maps.
• Introduced a suppression-ranking-based unsupervised training strategy to mitigate noise and emphasize foreground objects.

Main Results:

• The proposed DLUT demonstrated superior performance compared to state-of-the-art unsupervised trackers.
• Decoupled learning unlocked the potential of individual branches, leading to improved feature focus and learning.
• Independent adaptive modules generated more effective feature response maps for accurate object localization.

Conclusions:

• DLUT effectively overcomes the performance bottleneck caused by coupled branches in unsupervised object tracking.
• The novel training strategy successfully suppresses noise and enhances foreground object detection.
• The decoupled approach represents a significant advancement in unsupervised object tracking methodologies.