Two-stage learning to robust visual track via CNNs

Convolutional Neural Networks (CNN) are an alternative type of deep neural network that can be used to model local correlations and reduce translation variations, which have demonstrated great performance in some computer vision areas except the visual tracking due to the lack of training data. In this paper, we explore applying a two-stage learning CNN as a generic feature extractor offline pretrained with a large auxiliary dataset and then transfer its rich feature hierarchies to the robust visual tracking task. Instead of traditional neuron models in CNNs, we introduce a strategy to use ReLU for training acceleration. Empirical comparisons prove our CNN based tracker outperforms several state-of-the-art methods on an open tracking benchmark.