Object tracking with dynamic weighting Siamese network

Siamese-based models achieves outstanding performance in visual tracking. However, as the template is invariant during tracking, results in failing to capture the appearance changes over time, leading to incredible results. To tackle this problem, an object tracking method with dynamic weighting Siamese network has been proposed. First, we design a temporal attention module to obtain the time-sensitive features, which is learned by adaptively weighting the consecutive historical reference frames. Then, a spatial attention module is adopted to enhance the spatial information of the target, and promote the discriminative ability of the network. Finally, the Focal Loss is utilized to train the network offline, to improve the ability of handling the hard negative samples. We conducted the experiment on OTB2015 benchmark to evaluate the performance of the proposed method and other state-of-the-arts, comparison results demonstrate the good performance of our model.