Distractor-Aware Deep Regression for Visual Tracking.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Regression trackers are popular for visual object tracking due to performance and ease of implementation.
• Extreme imbalanced sample distribution in real-world data often compromises tracker robustness and accuracy.

Purpose of the Study:

• To develop a novel distractor-aware loss function to address sample imbalance in regression trackers.
• To enhance the accuracy and robustness of visual object tracking algorithms.

Main Methods:

• Proposed a distractor-aware loss function to balance training samples by emphasizing relevant domains and penalizing background.
• Introduced a differentiable hierarchy-normalized concatenation connection for multi-layer abstraction exploitation.

Main Results:

• The proposed tracker demonstrated significantly improved performance compared to state-of-the-art approaches.
• Experiments were conducted on five challenging benchmark datasets: OTB-13, OTB-15, TC-128, UAV-123, and VOT17.

Conclusions:

• The novel loss function effectively mitigates issues arising from imbalanced training data in regression trackers.
• The proposed approach offers a promising advancement in visual object tracking technology.