LoongTrack: Exploring long-sequence modeling for visual tracking.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Visual object tracking is a complex problem involving temporal and spatial data.
• Current transformer-based methods have quadratic complexity, limiting spatiotemporal cue extraction.
• Existing pipelines often overlook inherent frame information due to complex fusion modules.

Purpose of the Study:

• To develop a foundational visual object tracker with linear complexity.
• To exploit inherent spatiotemporal information within video frames more effectively.
• To enable efficient long sequence modeling for successive information flow.

Main Methods:

• Proposed causal consistent scanning to address the causal nature of tracking.
• Developed a simple tracking pipeline for long sequence modeling (time series and resolution).
• Systematically designed selective scan patterns considering temporal and spatial dimensions.

Main Results:

• Achieved encouraging results on multiple public datasets.
• Demonstrated reduced parameter count and training memory consumption.
• Validated the effectiveness of the proposed causal consistent scanning and selective scan patterns.

Conclusions:

• The proposed foundational tracker offers an efficient alternative to complex transformer-based methods.
• Exploiting inherent frame information and causal scanning enhances visual object tracking.
• This work provides inspiration for enriching spatiotemporal information in tracking tasks.