Good Features to Correlate for Visual Tracking.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Correlation filters (CF) excel in visual object tracking.
• Feature representation significantly impacts CF tracker performance.
• Current CF trackers often rely on pre-trained classification networks, limiting adaptability.

Purpose of the Study:

• To formulate and address the challenge of learning deep fully convolutional features specifically for CF-based visual tracking.
• To develop a flexible learning framework that alleviates dependency on classification-trained networks.

Main Methods:

• Proposed a novel and efficient backpropagation algorithm tailored to the CF tracking loss function.
• Fine-tuned convolutional layers of a state-of-the-art deep network for custom feature learning.
• Integrated the learned deep features into a top-performing CF tracker.

Main Results:

• Achieved an 18% increase in expected average overlap (EAO).
• Reduced tracking failures by 25%.
• Demonstrated superior performance over state-of-the-art methods on OTB-2013 and OTB-2015 datasets.

Conclusions:

• The proposed custom deep feature learning framework enhances CF-based visual object tracking.
• Fine-tuning networks offers a flexible and effective approach for improving tracking robustness and accuracy.