Visibility-aware Multi-teacher Knowledge Distillation for Siamese Tracker
Yiding Li, Cheng Chun Tang, Tsubasa Minematsu, Atsushi Shimada · 2024
In recent years, Siamese network-based trackers brought new vitality to the visual object tracking field. However, tracking tasks have always been troubled by complex scenarios. As Siamese trackers become more powerful, the performance bottlenecks caused by complex scenarios become more and more non-negligible. Occlusion is the most common and challenging complex scenario that can easily cause tracking failures. Some high-quality tracking databases provide visible ratio labels to describe occlusion in more detail. In addition, high-performance Siamese trackers can not run efficiently on resource-limited devices due to their high memory cost and complexity. To address these issues, we propose an Adaptive Multi-teacher Knowledge Distillation (AMKD) model to distill lightweight tracker, which is fast and achieves satisfactory performance in low visible ratios scenarios. In AMKD, we adopt the teacher model to transfer adequate knowledge to student. Furthermore, to extract visibility-based knowledge from visible ratios labeled data and transfer it to student efficiently, we introduced assistant teachers which are customed to overcome low visible ratios scenarios. For multiple assistant teachers transfer knowledge to student more efficiently and effectively, the AMKD is equipped with an Adaptive Selection Mechanism (ASM). Experiments of several Siamese trackers on high-quality dataset GOT-10K demonstrated the effectiveness of our method. Moreover, the AMKD distilled student achieve 9 times of compression rates and 6 times of speed up reach 181 FPS while improving accuracy in low visible ratios scenarios and obtaining favorable overall performance.