CNN feature or handcrafted feature in DCF object tracker?
Yan Zhou, Hongwei Guo, Dongli Wang · Engineering Science and Technology an International Journal · 2020
In the DCF framework, a crucial component except DCF trackers is the adopted feature map, especially when we need the tracker to be applied in practical. The powerful feature map from CNN achieves an outstanding performance when compared to the handcrafted feature (HOG, Color-Name), however, it makes the tracker slower and could not meet the need in the real-time scene. In this paper, we visualize the respective feature maps, filter, and location score from CNN, HOG and ColorName, to make a comparison. We also dig into the detail of the target label, for balance the accuracy and robustness of the tracker. Experiments on OTB-2015 show the performance loss from handcrafted feature is acceptable for the real-time application.