Adaptive online performance evaluation of video trackers

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Evaluating video tracking algorithms often requires ground-truth data, limiting applicability.
• Assessing tracker performance during failures and recoveries remains a significant challenge.

Purpose of the Study:

• To propose an adaptive framework for estimating video tracking quality without ground-truth data.
• To enable evaluation of trackers across long sequences with multiple failures and recoveries.

Main Methods:

• A two-stage framework: estimating tracker condition (target loss) and measuring track quality during success.
• Utilizing tracker uncertainty to identify successful tracking.
• Employing the time-reversibility constraint to determine track recovery from errors.

Main Results:

• Demonstrated effectiveness and robustness on a particle filter tracker with a heterogeneous dataset.
• The framework successfully handles tracking challenges like occlusions, illumination changes, and clutter.
• Improved state-of-the-art performance in sequences with multiple tracking errors and recoveries.

Conclusions:

• The proposed framework offers a robust method for evaluating video tracking algorithms without ground-truth data.
• It significantly enhances the assessment of tracking quality, particularly in complex scenarios with intermittent failures.
• This approach advances the field by providing a more comprehensive and practical evaluation tool for video tracking.