Attentional Correlation Filter Network for Adaptive Visual Tracking <Supplementary Material>

To show the effect of the parameters used in the Attentional Correlation Filter Network (ACFN), two additional experiments were conducted. In the first experiment, we varied the number of selected tracking modules (Na) in order to validate the robustness of the attentional mechanism, as shown in Fig. 2 (a). For this experiment, the number of tracking modules with high predicted validation scores (k) was fixed to 13. The result shows that the robustness of the tracker reduces dramatically when Na was too small, which was due to the insufficient variety of the considered properties in this case. The robustness also decreased with large Na, which meant that adding the tracking modules without considering the dynamic properties disturbed the robustness of the tracker. In the second experiment which is depicted in Fig. 2 (b), k was varied, while Na was fixed to 52. The performance dropped when k was set to small values, which was due to the insufficient number of tracking modules with a high predicted validation score. When k is too big, the performance of the tracker decreased because the prediction errors caused by inactive tracking modules were accumulated over time.