Geometric Hypergraph Learning for Visual Tracking.

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Graph-based methods are common in visual tracking, but often rely on pairwise relations.
• Existing methods struggle with target deformation and occlusion due to limited use of intrinsic structure.
• Pairwise affinities can be disturbed by errors, impacting tracking accuracy.

Purpose of the Study:

• To propose a robust visual tracking method using geometric hypergraph learning.
• To exploit high-order geometric relations among multiple correspondences for improved tracking.
• To enhance tracker scalability and robustness through confidence-aware sampling.

Main Methods:

• Formulating visual tracking as a mode-seeking problem on a hypergraph.
• Representing correspondence hypotheses as vertices and high-order relations as hyperedges.
• Developing a confidence-aware sampling method for constructing the geometric hypergraph.

Main Results:

• The proposed method effectively utilizes high-order geometric relations.
• Confidence-aware sampling enhances hypergraph robustness and scalability.
• Experiments show favorable performance against existing trackers on benchmark datasets.

Conclusions:

• Geometric hypergraph learning offers a powerful approach for visual tracking.
• Exploiting high-order relations significantly improves tracking performance under challenging conditions.
• The method demonstrates superior robustness and scalability compared to traditional graph-based trackers.