SPD-Updater: Symmetric positive definite manifold geometry based temporal updating for visual object tracking

Area of Science:

• Computer Vision
• Machine Learning
• Geometric Deep Learning

Background:

• Visual object tracking relies on template-based frameworks, often limited by fixed templates during appearance changes or occlusion.
• Existing dynamic template methods use Euclidean space for reliability scores, which can be unreliable in high-dimensional feature spaces.

Purpose of the Study:

• To develop a more robust online adaptation method for visual object tracking.
• To address the limitations of Euclidean metrics in high-dimensional feature spaces for dynamic template reliability assessment.

Main Methods:

• Exploited the geometric representation capacity of the Symmetric Positive Definite (SPD) manifold.
• Designed a novel score prediction module, termed SPD-Updater, for tracker updates.
• Utilized an SPD manifold metric for more accurate and stable dynamic template computation.

Main Results:

• The SPD-Updater demonstrated enhanced model capacity for handling complex tracking scenarios.
• Experimental validation on LaSOT, GOT-10k, TrackingNet, and UAV123 datasets confirmed the approach's effectiveness.
• The SPD metric proved beneficial for online tracking adaptation.

Conclusions:

• The proposed SPD-Updater significantly improves visual object tracking performance by leveraging SPD manifold geometry.
• The study highlights the superiority of manifold metrics over Euclidean metrics for dynamic template reliability in high-dimensional spaces.
• This work offers a promising direction for advancing robust and adaptive visual object tracking.