ScalableTrack: Scalable One-Stream Tracking via Alternating Learning.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Transformer-based one-stream trackers are prevalent for visual object tracking.
• Existing trackers have fixed computational dimensions, limiting context and global representation learning.
• This leads to reduced ability in distinguishing targets from backgrounds.

Purpose of the Study:

• To propose a new scalable one-stream tracking framework, ScalableTrack.
• To enhance object sensitivity and obtain discriminative global representations.
• To improve the ability of trackers to learn context clues and global representations.

Main Methods:

• Introduced a scalable one-stream tracking framework (ScalableTrack).
• Unified feature extraction and information integration via intrastage mutual guidance.
• Employed an alternating learning strategy to bridge interstage contextual cues.

Main Results:

• ScalableTrack demonstrated superior performance on eight challenging benchmarks.
• The method showed enhanced generalization and global representation capabilities.
• Outperformed state-of-the-art (SOTA) visual object tracking methods.

Conclusions:

• ScalableTrack effectively addresses limitations of fixed computational dimensions in one-stream trackers.
• The proposed framework achieves better object sensitivity and discriminative global representations.
• The alternating learning strategy prevents catastrophic forgetting and improves tracking accuracy.