TARA: Tracking with Aspect Ratio Adaptability

In visual object tracking (VOT), accurate and robust scale estimation of a target object is a challenging task. The discriminative correlation filter (DCF) is widely employed in VOT due to its high efficiency and accuracy. However, DCF based trackers do not have inherent scale adaptability. Most existing scale estimation methods for DCF based trackers cannot accommodate aspect ratio variation and thus result in inferior performance. In this paper, we propose to address the scale estimation problem and enable aspect ratio adaptability by utilizing a group of DCFs to localize the boundaries of the target object. Deep hierarchical convolutional features are exploited to improve the accuracy and robustness. The resulting system is named TARA: tracking with aspect ratio adaptability. Extensive empirical evaluation using the publicly available tracking benchmark datasets demonstrates that TARA can meet the demand of scale variation challenges and obtains favorable performance compared to state-of-the-art trackers.