Toward Robust Visual Object Tracking With Independent Target-Agnostic Detection and Effective Siamese Cross-Task...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Advanced Siamese networks excel in visual object tracking but struggle with severe appearance variations and lack cross-task synergy.
• Existing methods independently design classification and regression modules, hindering collaborative target localization.

Purpose of the Study:

• To develop a Siamese visual object tracking framework that addresses limitations in handling appearance variations and promotes cross-task interaction.
• To improve the robustness and accuracy of visual object tracking through enhanced target detection and unified multi-task learning.

Main Methods:

• Introduced a novel network with a target-agnostic object detection module to mitigate appearance variation issues.
• Developed a cross-task interaction module for consistent supervision of classification and regression branches.
• Implemented adaptive labels for more effective network training in multi-task learning.

Main Results:

• The proposed target-agnostic detection module effectively complements direct target inference.
• Cross-task interaction module demonstrated improved synergy between classification and regression branches.
• Experimental results on OTB100, UAV123, VOT2018, VOT2019, and LaSOT benchmarks show superior tracking performance compared to state-of-the-art methods.

Conclusions:

• The novel Siamese tracking framework with target-agnostic detection and cross-task interaction significantly enhances tracking performance.
• The proposed methods effectively address challenges posed by severe appearance variations and improve multi-task learning synergy.
• This work offers a more robust and accurate solution for visual object tracking applications.