Learning to Rank Proposals for Siamese Visual Tracking.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Siamese network trackers with region proposal networks (RPN) are popular for visual tracking, decomposing the task into classification and regression.
• Existing methods treat all training samples equally and rely solely on classification scores for target localization during inference.

Purpose of the Study:

• To develop a strategy for ranking training sample importance in Siamese trackers.
• To enhance classification optimization by focusing on more informative samples.
• To improve online tracking performance through a refined proposal selection criterion.

Main Methods:

• Propose a strategy to rank training sample importance, giving more weight to crucial samples.
• Introduce a lightweight ranking network to generate scores for region proposals based on Intersection over Union (IoU) with ground-truth.
• Combine classification and ranking scores as a new criterion for online proposal selection.

Main Results:

• The proposed method significantly boosts tracking performance.
• Achieves state-of-the-art tracking accuracy.
• Maintains real-time tracking speed.

Conclusions:

• The proposed sample weighting and ranking strategy is effective for Siamese trackers.
• Easily integrated into existing RPN-based Siamese networks in an end-to-end manner.
• Demonstrates superior performance across 10 diverse tracking benchmarks.