Thermal Prompting for Object Tracking
Wenjuan Li, Huixian Chen, Jiong Zhao, Juanjuan Li · Frontiers in artificial intelligence and applications · 2025
Most existing RGB-Thermal (RGB-T) object tracking methods adopt the multi-modal images to fine-tune the pre-trained RGB-based models. However, these fine-tuning-based methods require designing specific fine-tuning strategies for different downstream tasks, which limits their flexibility. In this paper, we propose a Thermal Prompt visual Tracking model (TPT) to introduce prompt-learning to RGB-T tracking. We first present an Image Enhancement Module (IEM) to enhance high-frequency information contained in both RGB and infrared images via Fast Fourier Transform. Then, we design a Prompt Generation Network (PGN) to generate effective mixed-modal prompts. Moreover, during the prompt-tuning phase, the parameters of the pre-trained RGB-based model are frozen, and only the PGN module is trained to learn the multi-modality-related prompts. Experimental results on two benchmark datasets demonstrate the effectiveness of the proposed prompt learning strategy and the superiority of the TPT framework.