Performance Analysis of GIoU Loss Function for Object Tracking

The loss function plays a foremost role in the domain of tracking, specifically in the calculation of loss from the predicted bounding box to actual bounding box. The IoU loss function is commonly used in tracking algorithms. Nevertheless, it faces challenges when optimizing non-overlapping scenarios and does not actively contribute to the gradient in such situations. To address this issue, GIoU introduces an area compensation regularization to the IoU [121]. This inclusion effectively mitigates the problem of vanishing gradients that arise from bounding boxes having distinct overlapping directions or non-overlapping regions. Thus GIoU loss function has been incorporated in SiamFC++ and then trained and evaluated on the benchmark dataset. Further, the obtained results reveal that the incorporation of GIoU with fine-tuning improves the tracking performance by 1.8% on the LaSOT dataset. However, it struggles to achieve similar performance on other datasets. This necessitates the incorporation and analysis of other loss functions for object tracking.