Learning Background-Aware Correlation Filters for Visual Tracking - Supplementary Material

Spatial size of training samples: We evaluated the performance of our tracker over a range of different spatial support sizes on the OTB50 dataset, as shown in Table 1. We set the spatial size of training samples to be N2 times bigger than the target, where N 2 [2, ..., 5]. This experiment shows that increasing the support size improves the overlap precision, since more background patches are used for learning the tracker. However, since the tracking speed is linearly related to the support size, runtime performance suffers. We set the spatial support of training samples to be five times bigger than that of the target, N = 5 to trade-off between the accuracy and speed of our tracker.