Robust discriminative tracking via landmark-based label propagation.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Object appearance changes during tracking, violating independent identically distributed assumptions.
• Discriminative trackers require extensive training data, which is impractical for real-time visual tracking.

Purpose of the Study:

• To develop a novel, non-parametric discriminative tracker for robust visual object tracking.
• To address the challenge of limited training samples in dynamic visual tracking scenarios.

Main Methods:

• Utilizes landmark-based label propagation (LLP) on an undirected graph representation of samples.
• Employs a local landmarks approximation for cross-similarity matrix computation.
• Incorporates a graph Laplacian regularizer for soft label prediction and label diffusion.
• Integrates soft label predictions into a Bayesian inference framework for tracking.

Main Results:

• The proposed LLP tracker effectively propagates limited initial labels to a large volume of unlabeled samples.
• The method explicitly considers the local geometrical structure of all samples.
• Evaluations on 51 challenging image sequences show superior performance compared to state-of-the-art methods.

Conclusions:

• Landmark-based label propagation offers a powerful, non-parametric solution for visual tracking.
• The LLP tracker demonstrates robustness in handling objects with continuously changing appearances.
• This approach advances the field of discriminative visual tracking with limited training data.