End-to-end deep metric network for visual tracking

In this paper, we propose an end-to-end deep metric network (DMN) for visual tracking, where any target can be accurately tracked given only a bounding box of the first frame. Our main motivation is to make the network learn to learn a deep distance metric by following the philosophy of one-shot learning. Instead of utilizing a hand-crafted distance metric like Euclidean distance, our DMN focuses on providing a learnable metric, which is more robust to appearance variations. Furthermore, we are the first to properly combine mean square errors and contrastive loss into a joint loss function for back-propagation. During online tracking, DMN firstly applies our instance initialization for obtaining sequence-specific information and then straightforwardly tracks the target without the help of box refinement, occlusion detection and online updating. The final tracking score considers both our DMN scalar output and the constrain of motion smoothness. Ablation analyses are carried out to validate the effectiveness of our proposed method. And experiments on the prevalent benchmarks show that our method can achieve a competitive performance when compared with some representative trackers, especially those existing metric learning-based algorithms.