HROM: Learning High-Resolution Representation and Object-Aware Masks for Visual Object Tracking.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Siamese networks are crucial for object tracking, relying on feature cross-correlation.
• Current Siamese networks use low-resolution features, limiting bounding box accuracy.
• Accurate feature representation is vital for high-performance visual tracking.

Purpose of the Study:

• To develop a novel high-resolution Siamese network for robust object tracking.
• To improve spatial precision and semantic richness of feature representations.
• To enhance target discriminability against distractors and background.

Main Methods:

• Proposed a high-resolution Siamese network with parallel high-to-low resolution convolution streams.
• Implemented repeated information exchange across resolutions to maintain high-resolution representations.
• Utilized multi-scale feature fusion, attention mechanisms, and deformable convolutions.

Main Results:

• Achieved state-of-the-art performance on OTB100, UAV123, VOT2018, and LaSOT benchmarks.
• Demonstrated real-time tracking capabilities, confirming efficiency and effectiveness.
• The novel representation is semantically richer and spatially more precise.

Conclusions:

• The proposed high-resolution Siamese network significantly advances object tracking performance.
• The method offers a robust and efficient solution for accurate bounding box estimation.
• Effective feature representation and adaptive refinement are key to superior tracking.