SUIT: Spatial-Spectral Union-Intersection Interaction Network for Hyperspectral Object Tracking.

Area of Science:

• Computer Vision
• Machine Learning
• Signal Processing

Background:

• Hyperspectral videos (HSVs) offer rich spatial-spectral-temporal data beneficial for object tracking.
• Existing tracking methods often neglect spectral information, limiting performance in complex scenarios like cluttered backgrounds or small objects.

Purpose of the Study:

• To investigate and leverage spectral interactions for enhanced hyperspectral video object tracking.
• To develop a tracking framework that effectively integrates spatial and spectral cues.

Main Methods:

• Utilized Transformers to model band-wise long-range spatial relationships.
• Modeled spectral interactions using the inclusion-exclusion principle to integrate shared and band-specific spatial cues.
• Introduced a spectral loss function to enforce material distribution alignment during training.

Main Results:

• The proposed method achieves state-of-the-art performance in hyperspectral video tracking.
• Demonstrated improved robustness to shape deformation and appearance variations.
• Successfully integrated spatial and spectral information for superior tracking accuracy.

Conclusions:

• The developed approach effectively utilizes spectral interactions for robust and accurate hyperspectral video tracking.
• The findings highlight the importance of considering spectral information in addition to spatial cues for advanced tracking applications.
• Open-sourced code and models facilitate reproducibility and further research.