Siamese network based target-aware pixel-wise cross-correlation for Aerial Tracking

Inhalt:
Recently, because Siamese network can achieve real-time speed while maintaining superior accuracy, it has received great attention in the visual target tracking community. However, when there are similar interference objects around, the tracking effect is poor or even the tracking fails. Considering that there is more interference information of the target in aerial tracking, we propose a target-aware module which can mask the background information of the template image. Objects based on aerial tracking are relatively small, we further replace the traditional cross-correlation with part-to-part matching. Evaluations on several popular benchmarks show that the proposed SiamTP tracker can attain state-of-the-art results.