Res2net based Siamese Network for Object Tracking via Efficient Multi-scale Attention and Border Region Reppoints*

Object tracking is an important and challenging task. The changes of the object itself and the complex background can affect tracking performance. Siamese networks based object tracking are widely adopted due to their advantages in efficient similarity measurement, end-to-end learning, and modular design. However, existing object tracking algorithms based on Siamese networks do not extract object features sufficiently, resulting in lower tracking accuracy. And excessive parameter settings adversely affect real-time performance. Therefore, we propose a new algorithm, SiamBRR, which can more accurately and quickly track objects. We first introduce the Res2Net residual network into the Siamese network framework as the backbone feature extraction network to fully extract features. Then, we use the EMA to enhance feature representation. Furthermore, our proposed border region reppoints accurately locate the object border while avoiding the need for extensive parameter settings. Finally, we conducted experiments on three challenging public datasets: VOT2018, VOT2019, and OTB100. The experimental results illustrate that the proposed SiamBRR outperforms other advanced trackers in tracking accuracy.