Adaptive token division-based transformer for visual object tracking

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Standard one-stream trackers utilize all search tokens for template interaction across encoder layers.
• Search areas often contain distractors, leading to misidentification and reduced tracking accuracy.

Purpose of the Study:

• To propose a transformer-based visual object tracking framework with adaptive token division.
• To enhance the model's capability to differentiate between target objects and background clutter.

Main Methods:

• Developed a transformer-based tracking framework with an encoder-decoder structure.
• Introduced an adaptive token division module for optimized search and template token interaction.
• Implemented an attention masking strategy and Gumbel-Softmax technique for efficient, end-to-end optimization.

Main Results:

• The adaptive token division module improved the model's discrimination between objects and background.
• Experimental results on six benchmarks demonstrated the effectiveness of the proposed method.
• The framework achieved higher tracking accuracy by mitigating interference from distractors.

Conclusions:

• The proposed adaptive token division framework effectively addresses challenges in visual object tracking.
• The method offers a robust solution for accurate object tracking in complex visual scenes.
• This approach enhances the reliability and performance of transformer-based visual object trackers.