SiamFDA: feature dynamic activation siamese network for visual tracking.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Current Siamese network-based visual tracking algorithms often overlook global spatial information.
• This limitation leads to reduced robustness, particularly in scenarios with unreliable object regions.

Purpose of the Study:

• To introduce a novel anchor-free visual tracking framework, the Feature Dynamic Activation Siamese Network (SiamFDA).
• To address the limitations of existing methods by incorporating global spatial information and enhancing robustness.

Main Methods:

• Developed SiamFDA, a framework that captures long-range dependencies between distant pixels.
• Implemented a hierarchical feature selector for adaptive feature activation across different layers.
• Introduced an adaptive sample label assignment method to optimize training.

Main Results:

• SiamFDA demonstrated superior performance compared to state-of-the-art trackers across six benchmark datasets (VOT-2018, VOT-2019, GOT10k, LaSOT, OTB-2015, OTB-2013).
• The framework achieves a real-time processing speed of 40 frames per second.
• Achieved improved robustness in challenging tracking scenarios.

Conclusions:

• SiamFDA offers a significant advancement in anchor-free visual tracking.
• The proposed methods effectively leverage global spatial information and adaptive feature selection for enhanced performance.
• SiamFDA provides a robust and efficient solution for real-time visual object tracking.