Adaptive Part Mining for Robust Visual Tracking.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Visual tracking is crucial for estimating object states in videos.
• Drastic appearance changes and deformations pose significant challenges for existing trackers.
• Current methods often use fixed, coarse part divisions, limiting adaptability to diverse objects and deformations.

Purpose of the Study:

• To develop a robust visual tracking method that overcomes limitations of fixed part-based approaches.
• To introduce an adaptive part mining strategy for improved object representation and tracking accuracy.
• To enhance tracker performance in scenarios with significant appearance variations and distractors.

Main Methods:

• A transformer-based architecture comprising an object representation encoder, an adaptive part mining decoder, and an object state estimation decoder.
• Utilizing multiple part prototypes and cross-attention mechanisms for adaptive part capture.
• Implementing novel strategies within the state estimation decoder to manage appearance variations and distractors.

Main Results:

• The proposed adaptive part mining tracker (APMT) demonstrates robust performance in visual tracking tasks.
• Achieved promising results with high frames per second (FPS), indicating computational efficiency.
• Ranked first place in the VOT-ST2022 challenge, validating its effectiveness.

Conclusions:

• The APMT effectively addresses challenges in visual tracking, particularly appearance variations and arbitrary object deformations.
• The adaptive part mining decoder significantly improves the ability to capture relevant object parts.
• The tracker offers a robust and efficient solution for real-world visual tracking applications.