Refocus the Attention for Parameter-Efficient Thermal Infrared Object Tracking

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Deep learning models for thermal infrared (TIR) tracking face challenges due to limited training data.
• Full fine-tuning (FFT) of RGB models is common but inefficient and risks representation collapse.
• Existing parameter-efficient fine-tuning (PEFT) methods lack task-guided top-down attention.

Purpose of the Study:

• To introduce ReFocus, a new PEFT method for adapting RGB foundation models to TIR tracking.
• To enable task-guided top-down attention for improved representation learning in TIR tracking.
• To achieve state-of-the-art (SOTA) performance in TIR tracking with enhanced training efficiency.

Main Methods:

• ReFocus employs a top-down attention mechanism guided by high-level task-specific signals.
• The method freezes the entire foundation model, training only query-guided feature selection and top-down blocks.
• It adapts pretrained RGB models for downstream TIR tracking tasks.

Main Results:

• ReFocus achieves SOTA performance on five TIR tracking benchmarks.
• The method demonstrates significant performance improvements for foundation trackers in TIR tracking.
• Ablation studies confirm ReFocus's effectiveness, adaptability to lighter models, and different tracking frameworks.

Conclusions:

• ReFocus offers a superior alternative to FFT and bottom-up PEFT methods for TIR tracking.
• The approach achieves comparable or better performance with fewer training parameters and improved learning stability.
• ReFocus effectively leverages task-specific signals for efficient and high-performing TIR tracking.