Seeing Clear before Visual Tracking

In this paper, we propose a two-stages visual tracking method mainly based on two branches including image deblurring and visual tracking. Our main motivation is to achieve the robust visual tracking when the tracker is suffering fast motion blur. Firstly, we present the hierarchical model based on Spatial Pyramid Matching that performs the fine-to-coarse deblurring and exploits localized-to-coarse operations. After achieving the deblurred images, the proposed method use transformer framework with spatial and channel attention for extracting features in order to obtain the spatial and channel features simultaneously to obtain the fast visual tracking with the balance of accuracy and robustness. We first train the one-stage deblurring network in the dataset of Gopro. Then, we train the second stage visusal tracking branch. Lastly, we conduct extensive ablation studies to demonstrate the effectiveness of the proposed tracker, which obtains currently the outperforming results on large tracking benchmarks, we also validate the effectiveness of our method against the fast motion blurring.