Attention-based net for updating tracking template

During object tracking process, if only the first frame is used as the matching template, changes about the target state will often lead to poor tracking results or even tracking failure of the classic Siamese tracker. To deal with this issue, UpdateNet uses the first frame as template, and regularly updates the template with combination of the previous accumulated template and the current predicted template. However, the combining of template tends to bring in background information which may pollute the template representation. For the purpose of obtaining accurate template and timely sensing the change of target, this article introduces the Squeeze-and-Excitation channel attention and selective mechanism to UpdateNet. The channel attention mechanism can sort the template information spliced by channels by adjusting the weight to highlight important information. The confidence score of the tracking predicted result of the Siamese network is used to determine whether the corresponding frame should participate in template accumulation, and a threshold is set to exclude severely contaminated predicted templates. The article also uses a more detailed parameter adjustment method to enable UpdateNet to achieve convergence faster and be more adaptive. We apply the improved UpdateNet into the DaSiamRPN tracker, and evaluations on the VOT2016 and VOT2018 datasets show that our methods can effectively improve the performance of UpdateNet.