Does anybody know what the best loss function is for a highly imbalanced image dataset? I know that Focal Loss and Weighted Cross-Entropy are common solutions, but I can't find clear guidelines on when to choose one over the other. Any help here would be appreciated!

As I recall, you want FL if you handle highly imbalanced datasets. You are reducing the weight of well-classified examples. This is especially useful in object detection. Your situation with object detection is one where background examples far outnumber foreground objects, right?