Nocal-Siam: Refining Visual Features and Response With Advanced Non-Local Blocks for Real-Time Siamese Tracking

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Siamese trackers are popular for visual tracking due to their balance of accuracy and speed.
• However, their performance degrades when faced with distracters in the search region, leading to unreliable tracking results.

Purpose of the Study:

• To improve the robustness and accuracy of Siamese trackers by addressing their sensitivity to distracters.
• To introduce novel non-local blocks that leverage long-range dependencies for enhanced feature learning and target localization.

Main Methods:

• Proposed Nocal-Siam, integrating two novel non-local blocks: target-aware (T-Nocal) and location-aware (L-Nocal).
• T-Nocal refines features by learning target-guided weights, suppressing distracters and enhancing inter-branch feature interplay.
• L-Nocal associates response maps to prevent diverse candidate positions in subsequent frames.

Main Results:

• Nocal-Siam demonstrated superior performance compared to existing Siamese trackers.
• The proposed non-local blocks effectively mitigated the impact of distracters.
• Improved accuracy and reliability in target localization were observed across multiple benchmarks.

Conclusions:

• Nocal-Siam offers a significant advancement in Siamese tracking by effectively handling distracters through novel non-local attention mechanisms.
• The method shows strong potential for real-world applications requiring robust and accurate visual object tracking.