Efficient joint model learning, segmentation and model updating for visual tracking.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Tracking-by-segmentation is crucial for visual tracking, especially with appearance changes like deformation and occlusion.
• Existing methods often suffer from unreliable discriminative models due to inaccurate pseudo-labels and lack of spatial-temporal constraints.
• This leads to potential tracking failures in challenging scenarios.

Purpose of the Study:

• To improve the reliability and accuracy of tracking-by-segmentation algorithms.
• To address the limitations of current discriminative models in handling appearance variations.
• To develop a more robust visual tracking framework.

Main Methods:

• Integrated the objective function of model training directly into the target segmentation optimization framework.
• Constrained the discriminative model with spatial and temporal information during optimization.
• Introduced a supervision switch mechanism to detect and handle erroneous pseudo-labels by switching to a semi-supervised setting.

Main Results:

• Achieved more accurate target likelihoods for improved part labeling.
• Generated more reliable pseudo-labels for enhanced model learning.
• Demonstrated significant effectiveness on OTB2013, OTB2015, and TC-128 benchmarks.

Conclusions:

• The proposed method significantly enhances visual tracking performance by improving model training and segmentation optimization.
• The integration of spatial-temporal constraints and the supervision switch mechanism lead to more robust tracking.
• The approach effectively handles severe appearance changes and data distribution shifts.