The Importance of Feature Representation for Visual Tracking Systems with Discriminative Methods

Visual tracking has been a challenging problem in the field of computer vision due to a variety of appearance changes of target object, while it is a widely explored area, with many applications in human-computer interaction, surveillance and robotics. Recently thanks to a great progress made by machine learning researchers, more and more sophisticated techniques have been applied to visual tracking. In that case, the tracking task is easily translated into a binary classification problem. Based on this framework, In this paper we investigate the influence of different feature representations on the performance of a tracker by designing controlled experiments. Finally, we find that feature representation plays a crucial role in a visual tracking system. Additionally, although the complex model learning algorithm is the focus of many attentions and studies, our experiments indicate that a good feature representation is much more important than using a complex classification algorithm in a visual tracking system. We believe that our work will provide a fresh perspective for the research of visual tracking which can dramatically improve tracking performances.