Learning to Update for Object Tracking with Recurrent Meta-learner.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Model update is crucial for object tracking, typically treated as an online learning problem.
• Existing methods often rely on fixed update strategies like exponential moving average or stochastic gradient descent.

Purpose of the Study:

• To introduce a novel meta-learning framework for optimizing the model update process in object tracking.
• To develop a learned online learning algorithm that adapts and improves tracking models dynamically.

Main Methods:

• Formulating model update as a meta-learning problem, training an online learning algorithm on diverse offline video datasets.
• Designing a recurrent neural network (RNN)-based updater to process online training data and output updated target models.
• Integrating the learned updater into template-based and correlation filter-based trackers.

Main Results:

• The learned updater consistently improved the performance of base trackers.
• The proposed method achieved faster-than-real-time performance on GPU with a minimal memory footprint.
• Experiments showed superior performance compared to exponential moving average (EMA) and stochastic gradient descent (SGD) update baselines.
• The template-based tracker equipped with the learned updater reached state-of-the-art results among real-time GPU trackers.

Conclusions:

• Meta-learning offers a powerful paradigm for learning adaptive online learning algorithms in object tracking.
• The developed RNN-based learned updater provides significant improvements in both accuracy and efficiency for object tracking systems.
• This approach sets a new benchmark for real-time object tracking performance.