Adaptive Exploitation of Pre-trained Deep Convolutional Neural Networks for Robust Visual Tracking

Due to the automatic feature extraction procedure via multi-layer nonlinear\ntransformations, the deep learning-based visual trackers have recently achieved\ngreat success in challenging scenarios for visual tracking purposes. Although\nmany of those trackers utilize the feature maps from pre-trained convolutional\nneural networks (CNNs), the effects of selecting different models and\nexploiting various combinations of their feature maps are still not compared\ncompletely. To the best of our knowledge, all those methods use a fixed number\nof convolutional feature maps without considering the scene attributes (e.g.,\nocclusion, deformation, and fast motion) that might occur during tracking. As a\npre-requisition, this paper proposes adaptive discriminative correlation\nfilters (DCF) based on the methods that can exploit CNN models with different\ntopologies. First, the paper provides a comprehensive analysis of four commonly\nused CNN models to determine the best feature maps of each model. Second, with\nthe aid of analysis results as attribute dictionaries, adaptive exploitation of\ndeep features is proposed to improve the accuracy and robustness of visual\ntrackers regarding video characteristics. Third, the generalization of the\nproposed method is validated on various tracking datasets as well as CNN models\nwith similar architectures. Finally, extensive experimental results demonstrate\nthe effectiveness of the proposed adaptive method compared with\nstate-of-the-art visual tracking methods.\n