RGBT Tracking via Challenge-Based Appearance Disentanglement and Interaction

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Red-Green-Blue (RGB) and thermal imaging present unique challenges in target tracking.
• Robust target appearance representation is crucial for effective Red-Green-Blue and thermal (RGBT) tracking.

Purpose of the Study:

• To develop a novel approach for robust RGBT tracking by disentangling and interacting with target appearance representations.
• To address both modality-shared and modality-specific challenges in RGBT tracking.

Main Methods:

• Proposes a novel approach for target appearance representation disentanglement and interaction.
• Utilizes five challenge-based branches: three parameter-shared for modality-shared challenges and two parameter-independent for modality-specific challenges.
• Introduces a guidance interaction module for cross-modality feature transfer and an aggregation interaction module for combining representations.
• Designs a data generation strategy to create training data with diverse challenge attributes.

Main Results:

• The proposed tracker demonstrates superior performance compared to state-of-the-art methods.
• Achieves robust RGBT tracking across four benchmark datasets.
• Effectively models target appearance under various challenges, even with limited training data.

Conclusions:

• The novel approach significantly enhances RGBT tracking robustness and accuracy.
• The disentanglement and interaction strategy effectively leverages complementary modality information.
• The method offers a promising direction for future research in multi-modal visual tracking.