Siamese Tracking Network with Informative Enhanced Loss

Designing an effective and uniform framework to meliorate tracking performance is very meaningful and essential. However, existing methods merely focus on single positive and negative instances corresponding to the exemplar, thoroughly ignoring the effective information hidden in other instances. To tackle this issue, in this paper, we present an informative enhanced loss based Siamese tracking network. Specifically, we introduce an informative enhanced loss to enable the network to capture information from an overall perspective. In other words, we construct dense connections among instances and exemplar. More importantly, we prove that our proposed loss can be transformed into the logistic loss and the triplet loss under particular parameter settings. Experiments on prevalent benchmarks demonstrate that the Siamese frameworks trained with our proposed loss indeed obtain better tracking results than original ones, and achieve promising performance against several state-of-the-art trackers on the real-time challenge.