Multi-Task Deep Dual Correlation Filters for Visual Tracking.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Deep features from different network layers are often treated independently in visual tracking.
• This independent treatment limits the representational power of deep features for robust tracking.

Purpose of the Study:

• To propose a novel multi-task deep dual correlation filters (MDDCF) method for robust visual tracking.
• To address the limitation of independent deep feature utilization in existing visual tracking approaches.

Main Methods:

• A multi-task learning scheme is proposed to exploit multi-level deep features by treating target representation with individual features as a single task.
• The objective function of dual correlation filters is reformulated, and a new alternating optimization method is introduced for joint training of filters and network parameters.
• An effective object template update scheme is designed to capture target appearance variations.

Main Results:

• The proposed MDDCF tracker demonstrates favorable performance compared to state-of-the-art methods.
• Extensive experiments on seven benchmark datasets validate the effectiveness of the MDDCF approach.

Conclusions:

• The MDDCF method successfully integrates multi-level deep features through a multi-task learning framework.
• The proposed approach enhances visual tracking robustness by better exploring feature interdependencies and capturing appearance variations.