Generalized Hierarchical Co-Saliency Learning for Label-Efficient Tracking.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Visual object tracking is crucial for human-centered AI and human-machine interaction.
• Current state-of-the-art tracking methods require extensive, densely annotated video data for training.
• Manual video annotation is labor-intensive and time-consuming, posing a significant bottleneck.

Purpose of the Study:

• To develop a weakly supervised tracking method that reduces reliance on manual annotations.
• To achieve a balance between annotation costs and tracking performance.
• To enhance target representation in search images by leveraging unlabeled data.

Main Methods:

• Propose a weakly supervised tracking method based on co-saliency learning.
• Integrate the method into various tracking frameworks (CNN-based and Transformer-based).
• Utilize unlabeled frames to extract valuable visual information and generate co-salient attention maps.

Main Results:

• Achieve competitive performance compared to fully supervised trackers using only 3.33% of manual annotations.
• Demonstrate effectiveness across four general tracking benchmarks.
• Show superior performance on egocentric tracking, achieving 0.538 success on TREK-150, outperforming fully supervised trackers by 7.7%.

Conclusions:

• Weakly supervised tracking with co-saliency learning offers a cost-effective alternative to fully supervised methods.
• The proposed method effectively enhances target representation and tracking accuracy.
• This approach holds significant promise for reducing annotation burden in visual object tracking research.