An Accurate Refinement Pathway for Visual Tracking

Recently, in the field of visual object tracking, tracking algorithms combined with segmentation have achieved impressive results while using mask to label targets VOT2020 dataset. Most trackers get by increasing resolution through multiple upsampling modules and gradually summing features backbone network. However, this refinement pathway does not fully consider spatial information features, therefore, are perfect. In paper, cross-stage cross-resolution (CSCR) module is proposed for optimizing effect. This makes full use semantic high-level low-level fuses them skip connections achieve a very accurate Experiments were conducted on VOT dataset, experimental outperformed other excellent verified effectiveness algorithm paper.