Robust Template Adjustment Siamese Network for Object Visual Tracking.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Existing visual trackers often fail due to model degeneration, leading to drift and target loss.
• Appearance templates extracted from initial frames struggle with long-term target variations.

Purpose of the Study:

• To propose a novel visual tracking framework, the Template Adjustment Siamese Network (TA-Siam).
• To address model drift and target loss challenges in Siamese networks.
• To enhance tracking accuracy and robustness for long-term sequences.

Main Methods:

• Introduced TA-Siam, a framework with template adjustment and classification-regression subnetworks.
• The template adjustment module adaptively updates templates using subsequent frame features.
• Utilized rhombus labels to reduce classification errors and an effective regression loss for training.

Main Results:

• TA-Siam demonstrated state-of-the-art performance on challenging benchmarks (VOT2016, VOT2018, OTB50, OTB100, GOT-10K, LaSOT).
• Achieved a high tracking speed of 45 FPS.
• Effectively mitigated model drift and improved target localization accuracy.

Conclusions:

• The proposed TA-Siam framework significantly enhances visual tracking robustness and accuracy.
• Adaptive template adjustment is crucial for handling appearance variations in long-term tracking.
• TA-Siam offers a promising solution for real-time, high-performance visual object tracking.