The Multi-task Fully Convolutional Siamese Network with Correlation Filter Layer for Real-Time Visual Tracking

In recent years, the trackers based on the siamese network have achieved good performance on various benchmarks. However, most siamese trackers have difficulty in discriminating the similar objects and cannot benefit from the shallow features in the neural network. In this paper, we used three methods to solve the above problems. We use the VGGNet as the backbone of our networks instead of the most used AlexNet. We jointly train the correlation filter and the embedding similarity learning. The multi-task learning makes our tracker benefit from both the shallow and deep features in the neural network. We use the correlation filter as an attention module to make the tracker pay more attention to the object being tracked. Extensive experiments on benchmarks show that our approach yields 11.4% relative gain in OTB2015 and 33% relative gain in VOT2017 compared with the SiamFC. The proposed tracker can be real-time while achieving leading performance in OTB2013, OTB2015 and VOT2017.