Self-Supervised Deep Correlation Tracking.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Training feature extraction networks for object tracking typically demands extensive manual annotations, which is resource-intensive.
• Existing methods often rely on supervised learning, limiting their applicability due to annotation requirements.

Purpose of the Study:

• To develop an effective self-supervised learning-based tracker within a deep correlation framework.
• To reduce the reliance on manually annotated data for training feature extraction networks in video tracking.

Main Methods:

• Proposes a self-supervised deep correlation tracker (self-SDCT) utilizing a multi-cycle consistency loss for learning from adjacent video frames.
• Employs forward-backward prediction with pseudo-labels and a Siamese correlation tracking framework for training.
• Introduces a similarity dropout strategy and a cycle trajectory consistency loss to enhance the training process.

Main Results:

• The self-SDCT effectively learns feature extraction networks using self-supervised information from video sequences.
• The method demonstrates competitive tracking performance compared to state-of-the-art supervised and unsupervised trackers.
• Achieves strong results on standard video tracking evaluation benchmarks.

Conclusions:

• Self-supervised learning offers a viable alternative to manual annotation for training robust video trackers.
• The proposed self-SDCT framework provides an efficient and effective solution for object tracking tasks.
• This approach advances the field of unsupervised and self-supervised learning in computer vision applications.