SiamBAN: Target-Aware Tracking With Siamese Box Adaptive Network.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Object tracking faces challenges with variations in scale and aspect ratios.
• Existing methods often rely on multi-scale search or anchor-based schemes, limiting performance in complex scenarios.
• Anchor-free trackers have emerged but suffer from classification-regression inconsistency.

Purpose of the Study:

• To propose a novel, effective, and anchor-free object tracking method named Siamese Box Adaptive Network (SiamBAN).
• To address the scale variation and classification-regression inconsistency issues in object tracking.
• To achieve robust and efficient visual tracking without reliance on prior scale or anchor information.

Main Methods:

• Developed SiamBAN, a fully convolutional network that predicts target boxes in a per-pixel, anchor-free manner.
• Implemented a dual-task approach, separating objectiveness classification and bounding box regression.
• Introduced a no-prior box design for increased flexibility and a target-aware branch to resolve inconsistency.

Main Results:

• SiamBAN demonstrates promising performance across multiple challenging benchmarks, including VOT2018, VOT2019, OTB100, UAV123, LaSOT, and TrackingNet.
• The proposed tracker achieves a processing speed of 35 frames per second (FPS).
• The anchor-free, data-driven approach effectively handles scale variations and improves tracking accuracy.

Conclusions:

• SiamBAN offers a simple yet effective solution for object tracking, overcoming limitations of previous methods.
• The target-aware scale handling and dual-task design contribute to robust performance in diverse conditions.
• SiamBAN presents a flexible and efficient anchor-free tracking framework suitable for real-world applications.