Real-Time MDNet with Adaptive Feature Fusion for RGB-T Object Tracking
Shanlin Jiang · 2025
Given a video and its initial target, visual object tracking aims to locate the specific target in the following frames. Most trackers only utilize visible modality, leading to inferior performance on dark night. Current researchers have tried to involve thermal data to achieve satisfying performance in all conditions. The main issue on RGB-T object tracking is how to fuse those modalities. In this paper, we propose a real-time MDNet with adaptive feature fusion method. We introduce information entropy to obtain fusion weights, which can achieve adaptive fusion. Our method can achieve better performance on the public dataset against several fusion methods. Both quantitative and qualitative experiments validate the effectiveness of our method. Furthermore, our method can maintain real-time speed.