Towards Real-World Visual Tracking With Temporal Contexts.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Current visual trackers often fail in real-world scenarios due to limitations in handling ideal conditions.
• The prevalent tracking-by-detection paradigm overlooks crucial temporal context, limiting performance.
• Existing methods inadequately leverage temporal information within templates, failing to utilize consecutive frame data effectively.

Purpose of the Study:

• To propose a robust visual tracking framework, TCTrack++, designed for real-world conditions.
• To enhance the utilization of temporal contexts in visual tracking.
• To improve the performance and applicability of visual trackers in dynamic environments.

Main Methods:

• Developed a two-level framework (TCTrack) for efficient temporal context exploitation.
• Introduced TCTrack++ with an attention-based temporally adaptive convolution for feature enhancement.
• Implemented an adaptive temporal transformer for similarity map refinement and a curriculum learning strategy.
• Utilized online evaluation for assessing real-world performance.

Main Results:

• TCTrack++ demonstrated superior performance across 8 well-known benchmarks.
• The proposed attention-based convolution effectively integrates temporal information into spatial features.
• The adaptive temporal transformer significantly refines similarity maps using encoded temporal knowledge.
• Online evaluations confirmed TCTrack++'s readiness for real-world applications.

Conclusions:

• TCTrack++ offers a significant advancement in visual tracking, particularly for real-world applications.
• The framework's ability to exploit temporal contexts addresses key limitations of existing trackers.
• The proposed methods provide a strong foundation for future research in robust visual tracking.