Abstract
In recent years, several methods have been developed toutilize hierarchical features learned from a deep convolu-tional neural network (CNN) for visual tracking. However,as features from a certain CNN layer characterize an ob-ject of interest from only one aspect or one level, the per-formance of such trackers trained with features from onelayer (usually the second to last layer) can be further im-proved. In this paper, we propose a novel CNN based track-ing framework, which takes full advantage of features fromdifferent CNN layers and uses an adaptive Hedge methodto hedge several CNN based trackers into a single strongerone. Extensive experiments on a benchmark dataset of 100challenging image sequences demonstrate the effectivenessof the proposed algorithm compared to several state-of-theart trackers.