Small Object Tracking in Satellite Videos with Gaussian Wasserstein Distance

Tracking small objects is a major challenge in satellite video object tracking. The traditional metric, i.e. intersection over union (IoU) of bounding boxes, is sensitive to target scale, while minor location deviation can cause a significant drop. Especially for tiny objects occupying very limited pixels, the boxes may not overlap or be fully inclusive and thus the IoU metric fails to provide a consistent measure. This paper proposes a new method named GWD-SiamFC++, which introduces a novel bounding box metric based on the Gaussian Wasserstein distance (GWD) to the classical SiamFC++ framework to address this issue. GWD is a continuous and consistent metric, offering a more precise representation of pixel weights within bounding boxes and accurately distinguishing between non-overlapping or fully inclusive boxes. Besides, it features scale invariance and is insensitive to object size, making it more suitable for small object tracking. Furthermore, nonlinear transformations are devised based on the GWD concept and yield the training normalized Gaussian Wasserstein distance (TrGWD) and testing normalized Gaussian Wasserstein distance (TeGWD), which are integrated into the classical SiamFC++ framework for the training and the testing phases, respectively. Experimental results on the SatSOT dataset reveal that the proposed method attains a success rate of 48.2% and a precision rate of 70.6%, and maintains satisfying computational efficiency.