SiamOT: An Improved Siamese Network with Online Training for Visual Tracking

As a prevailing solution for visual tracking, Siamese networks manifest high performance via convolution neural networks and weight-sharing schemes. Most existing Siamese networks have adopted various offline training strategies to realize precise tracking by comparing the extracted target features with template features. However, their performances may degrade when dealing with unknown targets. The tracker is unable to learn background information through offline training, and it is susceptible to background interference, which finally leads to tracking failure. In this paper, we propose a twin-branch architecture (dubbed SiamOT) to mitigate the above problem in existing Siamese networks, wherein one branch is a classical Siamese network, and the other branch is an online training branch. Especially, the proposed online branch utilizes feature fusion and attention mechanism, which is able to capture and update both the target and the background information so as to refine the description of the target. Extensive experiments have been carried out on three mainstream benchmarks, along with an ablation study, to validate the effectiveness of SiamOT. It turns out that SiamOT achieves superior performance with stronger target discrimination abilities.