Recursive Least-Squares Estimator-Aided Online Learning for Visual Tracking.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Visual object tracking is often framed as a one-/few-shot learning problem.
• Recent methods use complex meta-learning for fast adaptation but require extensive offline training.
• This can lead to overfitting and high computational costs.

Purpose of the Study:

• To propose a simple, effective online learning approach for few-shot visual object tracking.
• To enable adaptation without offline training by incorporating a memory retention mechanism.
• To reduce the risk of catastrophic forgetting in continual learning scenarios.

Main Methods:

• A recursive least-squares estimator-aided online learning approach is proposed.
• An in-built memory retention mechanism allows the model to remember previously seen objects.
• Seen data can be safely removed from training, similar to continual learning.
• The approach was evaluated on multi-layer perceptrons (RT-MDNet) and convolutional neural networks (DiMP).

Main Results:

• Consistent improvements were observed on challenging tracking benchmarks.
• The method demonstrated effectiveness and efficiency in few-shot online adaptation.
• It successfully unveiled the power of modern online deep trackers with minimal extra computational cost.
• The memory retention mechanism proved effective in preventing catastrophic forgetting.

Conclusions:

• The proposed recursive least-squares estimator approach offers a powerful solution for few-shot online adaptation in visual object tracking.
• It eliminates the need for offline training while maintaining high performance and preventing knowledge forgetting.
• This method represents a significant advancement in efficient and effective online visual tracking.