Distilling Siamese Trackers with Attention Mask

In recent years, the introduction of Siamese network has brought new vitality to the object tracking community. However, high-performance Siamese trackers cannot run at a real-time speed on mobile devices due to their complex and huge model. Knowledge distillation is a common and effective model compression method, but it is difficult to be applied to the challenging task like object tracking. We find out the fundamental cause is that the imbalance between the foreground and background in the object tracking task, which aggravates the problem of insufficient feature extraction ability of small backbone. Therefore, we propose the attention mask distillation (AMD) to help the student tracker focus on the foreground area faster and more accurately. The attention mask can be easily obtained from the feature maps and brings fine-granularity to the traditional binary mask. The experimental results on OTB 100 and VOT20 18 show that our method enables the student tracker perform as well as the teacher tracker. At the same time, it's able to run on the CPU at a hyper-real-time of 66 fps and achieves nearly 9 times model compression ratio. Such low computational and storage costs make it possible to deploy high-performance trackers on resource-constrained platforms.