A Multi-Level Self-Distillation-Based Unified Tracker for Efficient RGB-T Tracking.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• RGB-Thermal (RGB-T) tracking combines RGB and thermal infrared (TIR) data to improve visual tracking robustness.
• Existing RGB-T trackers often use complex architectures, hindering efficiency.

Purpose of the Study:

• To propose a novel Multi-level Self-Distillation (MSD) framework for efficient RGB-T tracking.
• To adapt a one-stream RGB tracker for RGB-T settings without architectural modification or extra parameters.

Main Methods:

• Jointly processing RGB and TIR inputs through a shared backbone.
• Employing a combination of self-supervised (contrastive loss, modality-gap alignment loss) and supervised objectives (intermediate focal loss, modality-specific losses, fused tracking loss).

Main Results:

• Achieved state-of-the-art tracking accuracy on LasHeR, RGBT234, and GTOT benchmarks.
• Maintained the computational efficiency of the original RGB tracker.

Conclusions:

• Optimized training strategies can surpass complex architectural modifications in multi-modal tracking.
• The MSD framework offers significant practical advantages for real-world deployment.