IASA: An IoU-aware tracker with adaptive sample assignment.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Existing object trackers often use classification and regression branches but lack robust sample assignment strategies.
• Current methods struggle with reliable candidate location ranking using classification or combined scores.

Purpose of the Study:

• To propose an intersection over union (IoU)-aware tracker with adaptive sample assignment (IASA) for enhanced object tracking.
• To improve the definition and utilization of positive and negative training samples.
• To achieve more accurate candidate location ranking and robust tracking performance.

Main Methods:

• Developed an IoU-aware classification score for precise candidate location ranking.
• Introduced IoU-focal loss for training an anchor-free tracker to predict classification scores.
• Proposed an adaptive sample assignment (ASA) strategy based on sample IoU characteristics.
• Utilized a star-shaped box feature representation to refine classification features.

Main Results:

• The IASA tracker demonstrated state-of-the-art performance across seven public datasets.
• The IoU-aware classification score significantly improved the ranking of candidate tracking locations.
• The adaptive sample assignment strategy effectively distinguished positive and negative training samples.

Conclusions:

• The proposed IASA tracker, combining IoU-aware scoring and adaptive sample assignment, offers a significant advancement in object tracking.
• The method effectively addresses limitations in sample definition and candidate ranking in existing trackers.
• IASA provides a robust and accurate solution for object tracking tasks.