Online structural learning with dense samples and a weighting kernel

A great deal of visual tracking algorithms, especially the tracking-by-detection methods, have been reported in recent years. Among them the structural learning based have shown great performance. One major problem with online learning a classifier is the limited number of training samples. Meanwhile, the success of correlation filters reveals the importance of allowing dense sampling in the training process. In this paper we propose to boost the robustness and the efficiency of an online learned structural support vector machine (SVM). Specifically, we find training with dense samples could be very efficient by applying the Fourier techniques and careful implementations. Furthermore, we propose to use a weighting kernel to improve tracking performance and the performance gain does not come with a sacrifice in the efficiency. Actually, the weighting kernel is as efficient as the linear kernel. Finally, we show favorable results on the latest VOT challenge sequences. An extended experiment incorporating a 34 dimensional HOG feature representation into our method results in top3 performance on the VOT-TIR2016 dataset.