Multiple convolutional features in Siamese networks for object tracking

Siamese trackers demonstrated high performance in object tracking due to their balance between accuracy and speed. Unlike classification-based CNNs, deep similarity networks are specifically designed address the image problem thus inherently more appropriate for task. However, mainly use last convolutional layers analysis target search, which restricts performance. In this paper, we argue that using a single layer as feature representation is not an optimal choice framework. We present Multiple Features-Siamese Tracker (MFST), novel algorithm exploiting several hierarchical maps robust tracking. Since provide abstraction levels characterizing object, fusing features allows obtain richer efficient of target. Moreover, handle appearance variations by calibrating extracted from two different CNN models. Based on advanced representation, our method achieves accuracy, while outperforming standard siamese tracker benchmarks.The source code trained models available at https://github.com/zhenxili96/MFST .