Tracking With Saliency Region Transformer.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Transformers have advanced visual tracking but often suffer from decreased speed with increased model capacity.
• Massively redundant information in tracking sequences poses a challenge for efficient and accurate visual tracking.

Purpose of the Study:

• To develop an efficient and accurate visual tracker that addresses the speed-accuracy trade-off in Transformer-based models.
• To mitigate feature inconsistencies arising from the two-stage design in visual tracking.

Main Methods:

• Proposed the Saliency Region Tracker (SRTrack), a heuristic two-stage tracker with a lightweight initial stage and a saliency-based discriminative stage.
• Introduced an attention scaling factor to enhance model robustness and address feature extrapolation issues between training and inference.

Main Results:

• SRTrack achieved a state-of-the-art Area Under Curve (AUC) of 0.699 on the LaSOT benchmark.
• The tracker demonstrated high efficiency, running at 61 Frames Per Second (FPS) on LaSOT.
• Experiments on large benchmarks confirmed SRTrack's superior efficiency and accuracy.

Conclusions:

• SRTrack effectively balances accuracy and speed in visual tracking, outperforming existing methods.
• The proposed attention scaling factor improves model robustness and performance, making it suitable for challenging tracking scenarios.