Joint Correlation and Attention Based Feature Fusion Network for Accurate Visual Tracking.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Visual object tracking relies on feature fusion methods like correlation and attention mechanisms.
• Correlation-based methods excel at location but miss semantic context.
• Attention-based methods capture semantics but neglect object position.

Purpose of the Study:

• To propose a novel tracking framework, JCAT, that synergistically combines correlation and attention mechanisms.
• To leverage the complementary strengths of both feature fusion approaches for robust visual object tracking.

Main Methods:

• JCAT employs parallel correlation and attention branches to generate distinct position and semantic features.
• Fusion features are created by directly summing the location and semantic features.
• A segmentation network, enhanced by a memory bank and filtering mechanism, performs pixel-wise state estimation.

Main Results:

• The JCAT tracker demonstrates highly promising performance across eight challenging visual tracking benchmarks.
• JCAT establishes a new state-of-the-art result on the VOT2018 benchmark.
• The integrated approach effectively balances location and semantic information for superior tracking.

Conclusions:

• The proposed JCAT framework successfully integrates correlation and attention mechanisms for advanced visual object tracking.
• JCAT offers a robust solution for pixel-wise state estimation, outperforming existing methods.
• The framework's design addresses limitations of individual correlation and attention approaches, leading to state-of-the-art results.