Self-Supervised Tracking via Target-Aware Data Synthesis.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Deep learning-based visual tracking methods require large, high-quality annotated datasets for effective training.
• Manual data annotation is expensive, time-consuming, and a significant bottleneck in developing robust tracking systems.

Purpose of the Study:

• To develop a self-supervised (SS) learning framework for visual tracking to overcome data annotation limitations.
• To enable existing deep trackers to be trained without human annotation by synthesizing diverse training data.

Main Methods:

• Introduced a novel 'crop-transform-paste' operation to synthesize training data by simulating appearance variations and background interference.
• Developed a target-aware data-synthesis method that integrates seamlessly into existing SS learning frameworks.
• Adapted existing deep trackers for SS training using the synthesized data, requiring no algorithmic modifications.

Main Results:

• Achieved competitive performance compared to supervised learning methods, especially with limited annotations.
• Demonstrated improved robustness against challenges like object deformation, occlusion (OCC), and background clutter (BC).
• Outperformed state-of-the-art unsupervised tracking methods and enhanced the performance of established supervised trackers (SiamRPN++, DiMP, TransT).

Conclusions:

• Self-supervised learning, powered by target-aware data synthesis, offers a viable alternative to supervised methods in visual tracking.
• The proposed method effectively reduces reliance on manual annotation while improving tracking accuracy and robustness.
• This SS learning mechanism can be readily integrated into existing visual tracking frameworks to boost their performance.