Visual object tracking based on siamese network with a deeper and more powerful backbone
Jun Zhang, Xiao Tan · 2020
Siamese tracker has become an important branch of single object tracking. It balanced speed and accuracy on the tracking task. The network of feature extraction is the key to performance of Siamese tracker. However, the recent trackers used classic backbone networks, such as AlexNet and ResNet. But there is a gap between these networks and the modern deep networks in performance. In order to improve the performance of Siamese tracker, we use a deeper and powerful modern backbone network to redesign the framework of Siamese Tracker. In this work, we did some convincing experiments on authoritative datasets, such as OTB2015, VOT2017 and VOT2018, which are frequently used by other trackers. The evaluation results on these datasets show that our tracker has better performance and robustness.