An Improved STARK Tracker with Gaussian Distributed Focal Loss
Jingxing Zhu, Hui Tang, Li Chai · 2024
The occlusion and motion blur in datasets pose significant challenges in the field of object tracking. To enhance the tracking performance, this paper proposes an improved STARK tracker with Gaussian Distributed Focal Loss (GDFL), in which bounding boxes are characterized as probability density distribution. GDFL models the bounding box as a Gaussian distribution based on the target’s visibility, and dynamically evaluates loss between distributions, which is more suitable for tracking tasks involving occlusion and motion blur. The advantages of our proposed method are verified by extensive experimental results on the Object Tracking Benchmark (OTB) dataset.