SpectralTracker: Jointly High and Low-Frequency Modeling for Tracking

Recently, a considerable number of top-performing Transformer based trackers have been proposed. However, most of them mainly focus on utilizing low-frequency information from a spatial-spectral analysis perspective, limiting their performance in complicated scenes. To address this problem, we propose a spectral tracker that explores how to capture high and low-frequency information for robust tracking jointly. Specifically, we design a novel dual-spectral information extraction and aggregation module (DSM) consisting of a high and low-frequency branch to capture and combine complementary frequency information of a Transformer effectively. Firstly, we divide the local window in the high-frequency branch to focus on more fine-grained high-frequency information. Then, in the low-frequency branch, we apply AvgPooling with a low-pass effect on a Transformer to amplify its low-frequency information. Furthermore, we design a shared MLP strategy to polarize the dual-frequency branching to high and low-frequency information attention. Finally, we utilize an MLP to complementarily fuse high and low-frequency information for frequency domain modeling. Comprehensive experiments on five tracking benchmarks (i.e., GOT-10k, TrackingNet, LaSOT, UAV123 and TNL2K) show that our spectral tracker achieves better performance than the state-of-the-art trackers.