Tracker Fusion on VOT Challenge: How Does It Perform and What Can We Learn about Single Trackers?

Tracker fusion i.e. the fusion of the outputs of different tracking methods is an interesting new concept. Thus it should also be considered in the VOT challenges. In this paper we evaluate the performance of tracker fusion on the VOT2013 and VOT2014 datasets. Furthermore, we utilize the fusion concept to create novel fusion based measures for evaluating trackers. Fusion based evaluation is interesting as it does not evaluate trackers independently but in the context of all other trackers. It allows us for example to identify trackers that could despite poor average performance be interesting for research in object tracking. We found e.g. that all state-of-the-art trackers lack some strengths of a simple NCC tracker. Tracker fusion can exploit this and profit from an additional NCC tracker. We raise the question: Can this also be exploited in a more direct way i.e. can we e.g. combine NCC concepts with a state-of-the-art tracker?