7 May 2024 in IEEE Transactions on Geoscience and Remote Sensing
The scarce onboard computational resources and real-time demand restrict the deployment of aerial trackers with sophisticated structures and customized operators. Meanwhile, aerial trackers need updating modules to adapt to continuous appearance variations in real-world long-term (LT) scenarios. However, frequent updating will introduce noisy templates and lead to tracking drifts and efficiency drops. Therefore, in this work, we develop a lightweight and highly efficient Siamese tracker for LT onboard aerial tracking applications. First, we build a compact, plain, and deployment-friendly Siamese network based on re-parameterization (Rep) as the baseline short-term (ST) tracker. Furthermore, we propose a tracking-specific decoupled knowledge distillation (KD) guided by strict teachers to unleash the appearance representation potential of the feature extractor without extra inference cost. Specifically, before distillation, the teacher conducts qualification verification to avoid misguiding the student. Then, hard negative background regions are mined and decoupled with the target region, encouraging the student to focus more on similar distractors and informative areas. Finally, to realize efficient and high-confidence LT tracking, we design two extensions and incorporate them into the boosted ST tracker: an initial-template-driven template updater with a corresponding pair-generating strategy to alleviate appearance pollution, and a confidence estimating branch to determine whether to update. Extensive results on large-scale drone benchmarks indicate that our proposed tracker significantly outperforms state-of-the-art (SOTA) aerial trackers. Real-world tests on our customized drone-captured LT dataset also validate its favorable practicability with a real-time speed of 44 fps on the Lynix KA200 chip.
9 October 2023 in IEEE Transactions on Multimedia
Although Siamese trackers have recently been gradually replaced by complicated and computationally expensive Transformer trackers, they are simpler and more applicable to real-world deployment. We believe that room for improvement still exists in the Siamese tracking framework and attribute the performance limitation to inadequate annotation, insufficient augmentation and suboptimal assignment, which heavily weakens the discriminative power. As Siamese methods directly inherit the assignment manner from object detection, they both face the imbalance between sparse annotated objects and dense background examples and between easy and hard negative examples. Moreover, existing augmentations and negative pairs are insufficient to simulate practical tracking ambiguity and failure cases. Nevertheless, work in the training vein is still overlooked. Therefore, we strive to yield a negative-driven training pipeline to unleash the potential of the Siamese framework without any extra inference cost. Specifically, 1) We devise strong negative augmentations based on random copy-paste to take full advantage of available annotations and generate more challenging tracking scenarios, especially negative examples. 2) We propose a semisupervised two-phase assignment that jointly utilizes existing annotations and model outputs to mine more appropriate and challenging negative examples. 3) We formulate a complementary reweighting loss by modifying the loss weight matrix to bridge subtasks and highlight the contributions of hard negative examples more smoothly. We choose several classic Siamese trackers to validate the pipeline effectiveness. After training, these trackers can gain, at most, a nearly 14% relative increase in performance, which is comparable to advanced Siamese trackers and even Transformer trackers. The experimental results indicate that the tracking-specific training pipeline is an efficient method for strengthening trackers and requires further development.