Dynamic Feature Fusion for MDNet-Based Tracking

Visual object tracking is an important topic in computer vision community, with potential applications, such as human-object interaction, autonomous driving and robotics. With the emerging of multi-modality data, thermal infrared (TIR) data are used to improve the performance and solve the challenging cases in extreme illumination, nighttime scenes. TIR data can provide shape information and be robust in low-light conditions. To fully leverage the complementary strengths of both modalities, we propose a dynamic feature fusion strategy and apply it on RT-MDNet. Our method introduces information entropy to evaluate the importance of both modalities, which are used to generate the fusion weight in a dynamic way. Our method is tested on the RGBT234 dataset, which shows a near real-time speed and achieve 0.502 of success rate. The experiments show satisfying performance on multi-modal feature fusion.