SiamAUDT: adaptive updating decision for online Siamese tracker

Most Siamese trackers use the first frame tracked object as a fixed template. To adapt to the changes in object appearances, some trackers have explored “how to update” templates. Meanwhile, to suppress negative samples from contaminating templates, some trackers utilize fixed thresholds to determine “updating or not”. To further improve the decision adaptability of “updating or not”, we first propose an online Siamese tracking framework with an adaptive updating decision (SiamAUDT). The decision module (AUD) can adaptively determine “updating or not” according to the confidence evaluation of the current tracking result. Second, we define a metric to assess the peak deviation degree (PDD) in the current response map. Then, we design the confidence evaluation method via the fluctuation amplitude of the PDD. The evaluation does not need any additional parameters or any training costs. Furthermore, the proposed framework can be easily migrated to existing tracking algorithms. Finally, we evaluated the effectiveness of our proposed SiamAUDT on several benchmarks and verified the generalization ability of the framework under several Siamese trackers.