i-Siam: Improving Siamese Tracker with Distractors Suppression and Long-Term Strategies

Recently, Siamese Network (SiamFC) has attracted much attention due to its fast tracking capability. Despite many improvements introduced, its accuracy is still far from human performance. In this work, we show that there are several design flaws in SiamFC tracker. In particular, the negative signals produced by SiamFC lead to noisy response map. In addition, background noises prevent SiamFC from extracting clean features from the template. To suppress these distractions, first we propose a negative signal suppression approach such that irrelevant features are deactivated. Secondly, we demonstrate that image-level suppression is also important in maximizing the tracking accuracy in addition to the existing feature-level suppression. With better detection sensitivity, we further propose a Diverse Multi-Template approach for appearance adaptation while reducing the risk of template drifting during long-term tracking. In our experiments, we conduct extensive ablation studies to demonstrate the effectiveness of the proposed components. Our tracker named improved Siamese (i-Siam) Tracker is able to achieve state-of-the-art results on UAV123, OTB-100, OxUvA, and TLP datasets compared to the existing trackers. Nonetheless, our tracker runs in real time, which is around 43 FPS.