Learning Precise Mask Representation for Siamese Visual Tracking.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Siamese network trackers excel in visual object tracking via similarity learning.
• Current bounding box formats struggle with non-rigid targets and background clutter, impacting accuracy.
• Existing methods are sensitive to distractors due to excessive background in bounding boxes.

Purpose of the Study:

• To develop a generic segmentation-assisted model for pixel-wise object tracking.
• To improve the accuracy and robustness of Siamese trackers for non-rigid targets.
• To enable seamless integration into existing Siamese tracking frameworks.

Main Methods:

• Proposed a novel segmentation-assisted model for learning binary mask representations.
• Implemented a multi-stage precise mask representation module with cascaded U-Net decoders.
• Introduced a saliency localization head using the Euclidean model for spatial constraints.

Main Results:

• The model effectively improves both anchor-based and anchor-free Siamese trackers.
• Achieved significant performance gains on five tracking benchmarks, including GOT-10k.
• Demonstrated increased Average Overlap (AO) scores for SiamRPN++ (5.2%) and SiamBAN (7.5%).
• Maintained high tracking speeds exceeding 60 FPS.

Conclusions:

• The proposed segmentation-assisted model offers a superior alternative to bounding box tracking.
• This generic approach enhances Siamese trackers' ability to handle complex deformations and reduce distractor sensitivity.
• The method significantly boosts tracking performance while preserving computational efficiency.