UncTrack: Reliable Visual Object Tracking With Uncertainty-Aware Prototype Memory Network.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Transformer-based trackers are dominant in object tracking due to their accuracy and efficiency.
• Existing methods often overlook target localization uncertainty, limiting performance in complex situations.
• Reliable target state prediction is crucial for robust object tracking.

Purpose of the Study:

• To propose UncTrack, an uncertainty-aware transformer-based tracker.
• To address the limitation of overlooking target localization uncertainty in current trackers.
• To enhance the robustness and accuracy of object tracking in challenging scenarios.

Main Methods:

• UncTrack predicts target localization uncertainty using an uncertainty-aware localization decoder (ULD).
• A prototype memory network (PMN) utilizes uncertainty information for reliable target state inference.
• High-confidence samples are used to update the prototype memory bank, improving template representation.

Main Results:

• UncTrack demonstrates superior performance compared to state-of-the-art object tracking methods.
• The incorporation of localization uncertainty leads to more reliable target state predictions.
• The method shows increased robustness against challenging appearance variations.

Conclusions:

• UncTrack effectively integrates localization uncertainty into transformer-based tracking.
• The proposed uncertainty-aware approach significantly enhances tracking performance and reliability.
• UncTrack represents a significant advancement in addressing limitations of deterministic tracking methods.