Cascaded Correlation Refinement for Robust Deep Tracking.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Deep learning trackers show high performance in visual tracking.
• Existing methods face challenges in robust target localization and reliable model updates.

Purpose of the Study:

• To propose a cascaded correlation refinement approach for enhancing the robustness of deep visual tracking.
• To improve accurate target localization and reliable model updates collaboratively.

Main Methods:

• Cascading multiple stages of correlation refinement for progressive target localization.
• Learning an accurate on-the-fly model using localized objects for reliable model updates.
• Introducing a measure for tracking failure identification and a look-back scheme for adaptive model updates.

Main Results:

• The proposed tracker demonstrates superior robustness compared to state-of-the-art methods.
• Achieved accurate target localization and reliable model updates through the cascaded refinement process.
• Validated performance across diverse benchmarks including OTB2013, OTB2015, VOT2016, VOT2018, UAV123, and GOT-10k.

Conclusions:

• The cascaded correlation refinement approach significantly enhances deep tracking robustness.
• The method effectively addresses challenges in target localization and model updating.
• The proposed tracker represents a state-of-the-art solution for robust visual tracking.