Learning Self-Corrective Network via Adaptive Self-Labeling and Dynamic NMS for High-Performance Long-Term Tracking.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Long-term object tracking is challenged by severe appearance variations and occlusions.
• Existing trackers often suffer from cumulative errors due to inaccurate online model updates, leading to drift.
• Robustness requires self-correction to judge tracking reliability and recapture targets after drift.

Purpose of the Study:

• To introduce a self-corrective network-based long-term tracker (SCLT) for improved robustness.
• To develop a tracking reliability evaluator (STRE) and a proposal postprocessor (SPPP) for target recapture.
• To address cumulative errors and drift issues in existing long-term tracking methods.

Main Methods:

• The STRE employs a modulation subnetwork and adaptive self-labeling for reliable tracking classification.
• The SPPP utilizes dynamic Non-Maximum Suppression (NMS) to recapture targets when drift is detected.
• The proposed dynamic NMS adaptively handles in-view and out-of-view scenarios for accurate object box selection.

Main Results:

• The SCLT demonstrates superior performance over state-of-the-art long-term trackers on challenging benchmark datasets (VOT2021LT, OxUvALT, TLP, LaSOT).
• The STRE and SPPP show good transportability and improve baseline tracker performance.
• Extensive evaluations confirm the effectiveness of the proposed self-corrective approach in all measures.

Conclusions:

• The SCLT effectively mitigates drift and improves long-term tracking accuracy through self-correction.
• The developed STRE and SPPP modules offer significant advancements in handling challenging tracking scenarios.
• The proposed methods provide a robust solution for real-world long-term object tracking applications.