Filter learning from deep descriptors of fully convolutional siamese network for tracking in videos

Abstract

In recent years, the advancement of Deep Learning has revolutionized many areas in Computer Vision, including Visual Object Tracking. A particular type of deep neural network, the Siamese Neural Network, brought the attention of Visual Object Tracking community. This neural network has a relatively low computational cost, and high e - cacy framework used to compare the similarity between objects. Nowadays, the scienti c community achieved remarkable success applying such frameworks in the tracking problem. However, the limitations this neural network impact negatively in its performance. We overcome this problem by obtaining a new descriptor for the reference object combining past descriptors outputted from the tracker. Speci cally, we propose a combination of the signal of descriptors in long and short term memory blocks, which represent the rst and the recent appearance of the object, respectively. A nal descriptor is composed of such memory blocks, and the tracker uses it as a reference. In particular, this work emphasized in the obtention of a method to compute an optimized lter bank through the usage of a genetic algorithm. The lter bank is then used to compute the short term memory output. According to experiments performed in the widely used OTB dataset, our proposal improves the baseline performance. The improvements for the area under the curve metrics are 7.4% and 3.0%, for precision and success plots, respectively, being comparable to the state-of-the-art methods.