Text this: Optimizing the loss function for bounding box regression through scale smoothing